Model Zoo¶
This is a list of pretrained ParlAI models. They are listed by task, or else in a pretraining section (at the end) when meant to be used as initialization for fine-tuning on a task.
Wikipedia models¶
Wikipedia Retriever (Used For Wizard Of Wikipedia)¶
Retrieval over Wikipedia dump, used for Wizard of Wikipedia.
Example invocation(s):
parlai interactive --model tfidf_retriever -mf zoo:wikipedia_full/tfidf_retriever/model
Enter Your Message: Yann LeCun
[candidate_scores]: [454.74038503 353.88863708 307.31353203 280.4501096 269.89960432]
[SparseTfidfRetrieverAgent]:
Yann LeCun (; born 1960) is a computer scientist with contributions in machine learning, computer vision, mobile robotics and computational neuroscience. He is well known for his work on optical character recognition and computer vision using convolutional neural networks (CNN), and is a founding father of convolutional nets. He is also one of the main creators of the DjVu image compression technology (together with Léon Bottou and Patrick Haffner). He co-developed the Lush programming language with Léon Bottou.
Yann LeCun was born near Paris, France, in 1960. He received a Diplôme d'Ingénieur from the Ecole Superieure d'Ingénieur en Electrotechnique et Electronique (ESIEE), Paris in 1983, and a PhD in Computer Science from Université Pierre et Marie Curie in 1987 during which he ...
of Science and Technology in Saudi Arabia because he was considered a terrorist in the country in view of his atheism.
In 2018 Yann LeCun picked a fight with a robot to support Facebook AI goals.
Wizard Of Wikipedia models¶
Wizard Of Wikipedia (End To End Generator)¶
End2End Generative model for Wizard of Wikipedia
Example invocation(s):
parlai display_model -t wizard_of_wikipedia:generator -mf zoo:wizard_of_wikipedia/end2end_generator/model -n 1 --display-ignore-fields knowledge_parsed
[chosen_topic]: Gardening
[knowledge]: no_passages_used __knowledge__ no_passages_used
Gardening __knowledge__ Gardening is the practice of growing and cultivating plants as part of horticulture.
Gardening __knowledge__ In gardens, ornamental plants are often grown for their flowers, foliage, or overall appearance; useful plants, such as root vegetables, leaf vegetables, fruits, and herbs, are grown for consumption, for use as dyes, or for medicinal or cosmetic use.
Gardening __knowledge__ Gardening is considered by many people to be a relaxing activity.
Gardening __knowledge__ Gardening ranges in scale from fruit orchards, to long boulevard plantings with one or more different types of shrubs, trees, and herbaceous plants, to residential yards including lawns and foundation plantings, to plants in large or small containers ...
there had been several other notable gardening magazines in circulation, including the "Gardeners' Chronicle" and "Gardens Illustrated", but these were tailored more for the professional gardener.
[title]: Gardening
[checked_sentence]: Gardening is considered by many people to be a relaxing activity.
[eval_labels_choice]: I live on a farm, we garden all year long, it is very relaxing.
[checked_sentence_parsed]: Gardening __knowledge__ Gardening is considered by many people to be a relaxing activity.
[WizTeacher]: Gardening
I like Gardening, even when I've only been doing it for a short time.
[eval_labels: I live on a farm, we garden all year long, it is very relaxing.]
[TorchAgent]: i love gardening , it is considered a relaxing activity .
Wizard Of Wikipedia (Full Dialogue Retrieval Model)¶
Full Dialogue Retrieval Model for Wizard of Wikipedia
Example invocation(s):
parlai display_model -t wizard_of_wikipedia -mf zoo:wizard_of_wikipedia/full_dialogue_retrieval_model/model -m projects:wizard_of_wikipedia:wizard_transformer_ranker --n-heads 6 --ffn-size 1200 --embeddings-scale False --delimiter ' __SOC__ ' --n-positions 1000 --legacy True
[chosen_topic]: Gardening
[knowledge]: Gardening Gardening is the practice of growing and cultivating plants as part of horticulture.
Gardening In gardens, ornamental plants are often grown for their flowers, foliage, or overall appearance; useful plants, such as root vegetables, leaf vegetables, fruits, and herbs, are grown for consumption, for use as dyes, or for medicinal or cosmetic use.
Gardening Gardening is considered by many people to be a relaxing activity.
Gardening Gardening ranges in scale from fruit orchards, to long boulevard plantings with one or more different types of shrubs, trees, and herbaceous plants, to residential yards including lawns and foundation plantings, to plants in large or small containers grown inside or outside.
Gardening Gardening may be very specialized, with only one type of plant grown, ...
there had been several other notable gardening magazines in circulation, including the "Gardeners' Chronicle" and "Gardens Illustrated", but these were tailored more for the professional gardener.
[title]: Gardening
[checked_sentence]: Gardening is considered by many people to be a relaxing activity.
[eval_labels_choice]: I live on a farm, we garden all year long, it is very relaxing.
[wizard_of_wikipedia]: Gardening
I like Gardening, even when I've only been doing it for a short time.
[label_candidates: OK what's the history?|Right, thats cool. I had no idea they still did the DVD thing, What is Netflix's highest rated show? do you know? |I will definitely check his first album out as he sounds interesting.|I don't know a whole lot about it. I was raised Catholic but don't practice anything now.|Well , this was a good conversation. |...and 95 more]
[eval_labels: I live on a farm, we garden all year long, it is very relaxing.]
[TorchAgent]: I live on a farm, we garden all year long, it is very relaxing.
parlai interactive -m projects:wizard_of_wikipedia:interactive_retrieval -t wizard_of_wikipedia
[ Your chosen topic is: Teapot ]
Enter Your Message: do you like tea?
[WizardRetrievalInteractiveAgent]: Yes! I only use teapots that have a little air hole in the lid. That prevents the spout from dripping or splashing when the tea is poured. Most teapots have this though.
Enter Your Message: what about kettles?
[WizardRetrievalInteractiveAgent]: I would think you could use them to heat any type of liquid! I use my teapots with a tea cosy. It's a thermal cover that helps keep the tea hot.
Enter Your Message: do you like earl grey?
[WizardRetrievalInteractiveAgent]: I think I'll try some Lipton, I love their green tea!
Imageseq2Seq Dodecadialogue Wizard Of Wikipedia Ft Model¶
Image Seq2Seq model trained on all DodecaDialogue tasks and fine-tuned on the Wizard of Wikipedia task
Example invocation(s):
parlai interactive -mf zoo:dodecadialogue/wizard_of_wikipedia_ft/model --inference beam --beam-size 10 --beam-min-length 10 --beam-block-ngram 3 --beam-context-block-ngram 3 -t wizard_of_wikipedia
Please choose one of the following topics by typing A, B, C, ..., etc. :
A: Denver Art Museum
B: Joke
C: Sport utility vehicle
D: Chocolate brownie
E: [NO TOPIC]
Enter Your Message: D
[ Your chosen topic is: Chocolate brownie ]
/private/home/kshuster/ParlAI/parlai/core/torch_ranker_agent.py:670: UserWarning: [ Executing eval mode with provided inline set of candidates ]
''.format(mode)
[ Using chosen sentence from Wikpedia ]: A chocolate brownie (commonly referred to as simply brownie) is a square, baked, chocolate dessert.
[ImageSeq2seq]: i love chocolate brownies . they ' re my favorite .
Enter Your Message: me too! do you ever make them?
[ Using chosen sentence from Wikpedia ]: Brownies are typically eaten by hand, often accompanied by milk, served warm with ice cream (a la mode), topped with whipped cream, or sprinkled with powdered sugar and fudge.
[ImageSeq2seq]: i don ' t make them , but i would love to .
parlai eval_model -mf zoo:dodecadialogue/wizard_of_wikipedia_ft/model -t wizard_of_wikipedia:Generator --prepend-gold-knowledge true
[ Finished evaluating tasks ['wizard_of_wikipedia:Generator'] using datatype valid ]
exs gpu_mem loss lr ppl token_acc total_train_updates tpb
3939 .3823 2.144 7.5e-06 8.532 .5348 22908 2852
Unlikelihood Wizard Of Wikipedia Context And Label Repetition Model¶
Dialogue model finetuned on Wizard of Wikipedia with context and label repetition unlikelihood
Example invocation(s):
python parlai/scripts/interactive.py -mf zoo:dialogue_unlikelihood/rep_wiki_ctxt_and_label/model -m projects.dialogue_unlikelihood.agents:RepetitionUnlikelihoodAgent
Enter Your Message: Hi.
[RepetitionUnlikelihood]: hi .
Unlikelihood Wizard Of Wikipedia Context Repetition Model¶
Dialogue model finetuned on Wizard of Wikipedia with context repetition unlikelihood
Example invocation(s):
python parlai/scripts/interactive.py -mf zoo:dialogue_unlikelihood/rep_wiki_ctxt/model -m projects.dialogue_unlikelihood.agents:RepetitionUnlikelihoodAgent
Enter Your Message: Hi.
[RepetitionUnlikelihood]: hi .
Unlikelihood Wizard Of Wikipedia Label Repetition Model¶
Dialogue model finetuned on Wizard of Wikipedia with label repetition unlikelihood
Example invocation(s):
python parlai/scripts/interactive.py -mf zoo:dialogue_unlikelihood/rep_wiki_label/model -m projects.dialogue_unlikelihood.agents:RepetitionUnlikelihoodAgent
Enter Your Message: Hi.
[RepetitionUnlikelihood]: hi .
Bart Fid Dpr Model¶
FiD model trained with a DPR Retriever and a BART-Large backbone seq2seq generator.
Example invocation(s):
parlai interactive -mf zoo:hallucination/bart_fid_dpr/model
Enter Your Message: Hey! What's your favorite Elvis song?
[Fid]: I love Elvis Presley, he is my favorite singer and songwriter. He was born in Memphis, TN and died in Memphis Tennessee.
parlai eval_model -mf zoo:hallucination/bart_fid_dpr/model -t wizard_of_wikipedia
Finished evaluating tasks ['wizard_of_wikipedia:WizardDialogKnowledge:random_split'] using datatype valid
accuracy bleu-1 bleu-2 bleu-3 bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem interdistinct-1 \
.0002539 .1772 .08068 .04245 .02717 76.62 78.62 55.81 0 0 .7098 3939 .2097 .04684 .1024
interdistinct-2 intradistinct-1 intradistinct-2 knowledge_f1 llen loss ltpb ltps ltrunc ltrunclen ppl \
.4421 .8833 .9822 .2172 23.61 2.687 23.61 16.76 0 0 14.68
rare_word_f1 rouge_1 rouge_2 rouge_L token_acc token_em tpb tps
.1547 .2647 .08819 .2244 .4348 0 102.2 72.57
Bart Fid Rag Dpr-Poly Model¶
FiD model trained with a RAG-DPR-Poly Retriever (DPR-Poly retrieval, trained in a RAG Setup) and a BART-Large backbone seq2seq generator.
Example invocation(s):
parlai interactive -mf zoo:hallucination/bart_fid_rag_dpr_poly/model
Enter Your Message: Hey! What's your favorite Elvis song?
[Fid]: I love Elvis Presley! He is my favorite singer, songwriter, actor, and producer.
parlai eval_model -mf zoo:hallucination/bart_fid_rag_dpr_poly/model -t wizard_of_wikipedia --num-examples 100
Finished evaluating tasks ['wizard_of_wikipedia:WizardDialogKnowledge:random_split'] using datatype valid
accuracy bleu-1 bleu-2 bleu-3 bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem interdistinct-1 \n.0002539 .1821 .09088 .05551 .03939 76.62 78.62 53.79 0 0 .6842 3939 .2154 .05641 .1056
interdistinct-2 intradistinct-1 intradistinct-2 knowledge_f1 llen loss ltpb ltps ltrunc ltrunclen ppl \n.4434 .8964 .9864 .2756 23.61 2.587 23.61 16.16 0 0 13.3
rare_word_f1 rouge_1 rouge_2 rouge_L token_acc token_em tpb tps
.1686 .2740 .1040 .2351 .4523 .0002539 102.2 69.95
Bart Fid Rag Model¶
FiD model trained with a RAG Retriever (DPR retrieval) and a BART-Large backbone seq2seq generator.
Example invocation(s):
parlai interactive -mf zoo:hallucination/bart_fid_rag/model
Enter Your Message: Hey! What's your favorite Elvis song?
[Fid]: I love Elvis Presley. I love his music. My favorite song is "My Way."
parlai eval_model -mf zoo:hallucination/bart_fid_rag/model -t wizard_of_wikipedia --num-examples 100
Finished evaluating tasks ['wizard_of_wikipedia:WizardDialogKnowledge:random_split'] using datatype valid
accuracy bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem knowledge_f1 llen loss lr ltpb ltps ltrunc ltrunclen ppl rare_word_f1 token_acc \
0 .04021 84.71 86.71 26.11 0 0 .3011 100 .2324 .03073 .2381 24.6 2.676 1e-05 24.6 7.407 0 0 14.53 .1678 .4301
token_em tpb tps
0 111.3 33.51
Bart Rag Dpr-Poly Model¶
RAG-Token model trained with a DPR Retriever and a BART-Large backbone seq2seq generator.
Example invocation(s):
parlai interactive -mf zoo:hallucination/bart_rag_dpr_poly/model
Enter Your Message: Hey! What's your favorite Elvis song?
[RAG]: I love Elvis Presley. He is my favorite singer of all time. Do you have a favorite song?
parlai eval_model -mf zoo:hallucination/bart_rag_dpr_poly/model -t wizard_of_wikipedia --num-examples 100
Finished evaluating tasks ['wizard_of_wikipedia:WizardDialogKnowledge:random_split'] using datatype valid
accuracy bleu-1 bleu-2 bleu-3 bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem interdistinct-1 \n.0005077 .1884 .09254 .05288 .03484 76.62 78.62 59.44 0 0 .7560 3939 .2230 .08013 .1053
interdistinct-2 intradistinct-1 intradistinct-2 knowledge_f1 llen loss ltpb ltps ltrunc ltrunclen ppl \n.4343 .9128 .9898 .2424 23.61 2.53 23.61 17.85 0 0 12.56
rare_word_f1 rouge_1 rouge_2 rouge_L token_acc token_em tpb tps
.1734 .2679 .09963 .2309 .4664 .0002539 102.2 77.29
Bart Rag Dpr Sequence Model¶
RAG-Sequence model trained with a DPR Retriever and a BART-Large backbone seq2seq generator.
Example invocation(s):
parlai interactive -mf zoo:hallucination/bart_rag_sequence/model
Enter Your Message: Hey! What's your favorite Elvis song?
[RAG]: My favorite Elvis Presley song is "Stuck on You". Do you have a favorite Elvis song?
parlai eval_model -mf zoo:hallucination/bart_rag_sequence/model -t wizard_of_wikipedia --num-examples 100
Finished evaluating tasks ['wizard_of_wikipedia:WizardDialogKnowledge:random_split'] using datatype valid
accuracy bleu-1 bleu-2 bleu-3 bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem interdistinct-1 \n.0007616 .1771 .08545 .04963 .03463 76.62 78.62 43.95 0 0 .5591 3939 .2102 .05834 .1125
interdistinct-2 intradistinct-1 intradistinct-2 knowledge_f1 llen loss ltpb ltps ltrunc ltrunclen ppl \n.4500 .9186 .9928 .2566 23.61 2.484 23.61 13.2 0 0 11.99
rare_word_f1 rouge_1 rouge_2 rouge_L token_acc token_em tpb tps
.1630 .2587 .09407 .2234 .4256 0 102.2 57.15
Bart Rag Dpr Token Model¶
RAG-Token model trained with a DPR Retriever and a BART-Large backbone seq2seq generator.
Example invocation(s):
parlai interactive -mf zoo:hallucination/bart_rag_token/model
Enter Your Message: Hey! What's your favorite Elvis song?
[RAG]: I love Elvis Presley. He is my favorite singer of all time. Do you have a favorite Elvis song?
parlai eval_model -mf zoo:hallucination/bart_rag_token/model -t wizard_of_wikipedia --num-examples 100
Finished evaluating tasks ['wizard_of_wikipedia:WizardDialogKnowledge:random_split'] using datatype valid
accuracy bleu-1 bleu-2 bleu-3 bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem interdistinct-1 \n.0002539 .1859 .08962 .04951 .03237 76.62 78.62 55.44 0 0 .7052 3939 .2200 .06284 .1012
interdistinct-2 intradistinct-1 intradistinct-2 knowledge_f1 llen loss ltpb ltps ltrunc ltrunclen ppl \n.4102 .9101 .9882 .2284 23.61 2.529 23.61 16.65 0 0 12.54
rare_word_f1 rouge_1 rouge_2 rouge_L token_acc token_em tpb tps
.1693 .2625 .09543 .2265 .4627 0 102.2 72.09
Bart Rag Dpr Turn Doc-Then-Turn Model¶
RAG-Turn Doc-Then-Turn model trained with a DPR Retriever and a BART-Large backbone seq2seq generator.
Example invocation(s):
parlai interactive -mf zoo:hallucination/bart_rag_turn_dtt/model
Enter Your Message: Hey! What's your favorite Elvis song?
[RAG]: I love Elvis Presley. He was a great singer and songwriter. Do you have a favorite song?
parlai eval_model -mf zoo:hallucination/bart_rag_turn_dtt/model -t wizard_of_wikipedia --num-examples 100
Finished evaluating tasks ['wizard_of_wikipedia:WizardDialogKnowledge:random_split'] using datatype valid
accuracy bleu-1 bleu-2 bleu-3 bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem interdistinct-1 \n.0005077 .1863 .0911 .05345 .03651 76.62 78.62 54.72 0 0 .6960 3939 .2202 .07843 .1032
interdistinct-2 intradistinct-1 intradistinct-2 knowledge_f1 llen loss ltpb ltps ltrunc ltrunclen ppl \n.4188 .9154 .9915 .2476 23.61 2.527 23.61 16.43 0 0 12.51
rare_word_f1 rouge_1 rouge_2 rouge_L token_acc token_em tpb tps
.1735 .2648 .09857 .2290 .4599 .0002539 102.2 71.15
Bart Rag Dpr Turn Doc-Only Model¶
RAG-Turn Doc-Then-Turn model trained with a DPR Retriever and a BART-Large backbone seq2seq generator.
Example invocation(s):
parlai interactive -mf zoo:hallucination/bart_rag_turn_do/model
Enter Your Message: Hey! What's your favorite Elvis song?
[RAG]: Hey! I love Elvis Presley. He is one of my favorite musicians. I love his voice.
parlai eval_model -mf zoo:hallucination/bart_rag_turn_do/model -t wizard_of_wikipedia --num-examples 100
Finished evaluating tasks ['wizard_of_wikipedia:WizardDialogKnowledge:random_split'] using datatype valid
accuracy bleu-1 bleu-2 bleu-3 bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem interdistinct-1 \n.0002539 .1844 .08733 .04707 .03028 76.62 78.62 51.39 0 0 .6537 3939 .2179 .09101 .1009
interdistinct-2 intradistinct-1 intradistinct-2 knowledge_f1 llen loss ltpb ltps ltrunc ltrunclen ppl \n.4128 .9113 .9890 .2253 23.61 2.658 23.61 15.43 0 0 14.26
rare_word_f1 rouge_1 rouge_2 rouge_L token_acc token_em tpb tps
.1656 .2582 .09185 .2228 .4461 0 102.2 66.83
Dropout Poly-Encoder¶
A Dropout Poly-encoder trained on the WoW knowledge selection task
Example invocation(s):
parlai tm -m rag --rag-retriever-type polyfaiss --query-model dropout_poly --poly-faiss-model-file zoo:hallucination/dropout_poly/model
TODO
Multiset Dpr Model¶
A DPR model pre-trained on Natural Questions and Trivia QA
Example invocation(s):
parlai tm -m rag --rag-retriever-type dpr --dpr-model-file zoo:hallucination/multiset_dpr/hf_bert_base.cp
TODO
Wikipedia Compressed Faiss Index¶
A compressed FAISS Index with embeddings of Wikipedia passages, generated by the Multiset DPR zoo model.
Example invocation(s):
parlai eval_model -mf zoo:hallucination/bart_rag_token/model --indexer-type compressed --path-to-index zoo:hallucination/wiki_index_compressed/compressed_pq -ne 100
TODO
Wikipedia Exact Faiss Index¶
A FAISS Index with embeddings of Wikipedia passages, generated by the Multiset DPR zoo model.
Example invocation(s):
parlai eval_model -mf zoo:hallucination/bart_rag_token/model --indexer-type exact --path-to-index zoo:hallucination/wiki_index_exact/exact -ne 100
TODO
Wikipedia Passages¶
A dump of Wikipedia articles split into 21m 100 word chunks from DPR (https://github.com/facebookresearch/DPR)
Example invocation(s):
parlai eval_model -mf zoo:hallucination/bart_rag_token/model --indexer-type exact --path-to-index zoo:hallucination/wiki_index_compressed/compressed --path-to-dpr-passages zoo:hallucination/wiki_passages/psgs_w100.tsv -ne 100
TODO
Wow Passages¶
A dump of Wikipedia articles split into ~3k paragraphs, comprising the subset of Wikipedia present in the WoW dataset.
Example invocation(s):
parlai eval_model -mf zoo:hallucination/bart_rag_token/model --indexer-type exact --path-to-index zoo:hallucination/wow_passages/exact --path-to-dpr-passages zoo:hallucination/wow_passages/wow_articles.paragraphs.tsv -ne 100
TODO
Wow Passages Compressed Index¶
A compressed FAISS Index of embeddings from ~3k paragraphs from Wikipedia present in the WoW dataset.
Example invocation(s):
parlai eval_model -mf zoo:hallucination/bart_rag_token/model --indexer-type compressed --path-to-index zoo:hallucination/wow_passages/compressed --path-to-dpr-passages zoo:hallucination/wow_passages/wow_articles.paragraphs.tsv -ne 100
TODO
Wow Passages Exact Index¶
A FAISS Index of embeddings from ~3k paragraphs from Wikipedia present in the WoW dataset.
Example invocation(s):
parlai eval_model -mf zoo:hallucination/bart_rag_token/model --indexer-type exact --path-to-index zoo:hallucination/wow_passages/exact --path-to-dpr-passages zoo:hallucination/wow_passages/wow_articles.paragraphs.tsv -ne 100
TODO
Light Dialog models¶
Light Bert-Biranker Dialogue Model¶
LIGHT Dialogue task, replicating the numbers from the paper.
Example invocation(s):
parlai eval_model -t light_dialog -mf zoo:light/biranker_dialogue/model
{'exs': 6623, 'accuracy': 0.7586, 'f1': 0.7802, 'hits@1': 0.759, 'hits@5': 0.965,'hits@10': 0.994, 'hits@100': 1.0, 'bleu': 0.7255, 'lr': 5e-05, 'total_train_updates': 15050,'examples': 6623, 'loss': 5307.0, 'mean_loss': 0.8013, 'mean_rank': 1.599, 'train_accuracy': 0}
Imageseq2Seq Dodecadialogue Light Dialogue Ft Model¶
Image Seq2Seq model trained on all DodecaDialogue tasks and fine-tuned on the LIGHT Dialogue task
Example invocation(s):
parlai interactive -mf zoo:dodecadialogue/light_dialog_ft/model --inference beam --beam-size 5 --beam-min-length 20 --beam-block-ngram 3 --beam-context-block-ngram 3
Enter Your Message: hi how's it going?
[ImageSeq2seq]: i ' m doing well . how about you ? what ' s going on in the world today ?
Enter Your Message: not much, wish it had some more epic battles!
[ImageSeq2seq]: me too . it ' s been so long since i ' ve seen a battle like this . do you have a favorite battle ?
parlai eval_model -mf zoo:dodecadialogue/light_dialog_ft/model -t light_dialog
[ Finished evaluating tasks ['light_dialog'] using datatype valid ]
exs gpu_mem loss lr ppl token_acc total_train_updates tpb
6623 .07002 2.927 7.5e-06 18.66 .3927 38068 20.81
Light Am I Me Or You Vanilla 128 Baseline¶
Vanilla 128-truncation baseline from the Am I Me or You project.
Example invocation(s):
parlai em -mf zoo:light_whoami/vanilla_128/model \
-t light_dialog:simple_multi:light_label=speech \
-ne 10
accuracy bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem llen loss lr ltpb ltps ltrunc ltrunclen ppl token_acc token_em tpb tps
0 1.067e-05 255.1 126.7 15.36 .9000 128.4 .1212 10 .1968 .1830 25.6 2.358 7e-10 25.6 3.104 0 0 10.57 .5078 0 152.3 18.46
parlai em --predictor-model-file zoo:light_whoami/rpa_reranker/model \
-t projects.light_whoami.task.agents:ResponseClassifierTeacher \
--model projects.light_whoami.agents.rpa_rerank:RPARerankAgent \
-mf zoo:light_whoami/vanilla_128/model -ne 10
accuracy bleu-4 character_accuracy ctpb ctps exps exs f1 gpu_mem lr ltpb ltps tpb tps
0 1.067e-05 1 126.7 14.52 .1146 10 .2065 .2802 7e-10 25.6 2.934 152.3 17.45
Light Am I Me Or You Vanilla 1024 Baseline¶
Vanilla 1024-truncation baseline from the Am I Me or You project.
Example invocation(s):
parlai em -mf zoo:light_whoami/vanilla_1024/model \
-t light_dialog:simple_multi:light_label=speech \
-ne 10
accuracy bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem llen loss lr ltpb ltps ltrunc ltrunclen ppl token_acc token_em tpb tps
0 1.152e-05 255.1 255.1 32.2 0 0 .1262 10 .1858 .1994 25.6 2.335 1e-09 25.6 3.231 0 0 10.33 .5078 0 280.7 35.43
parlai em --predictor-model-file zoo:light_whoami/rpa_reranker/model \
-t projects.light_whoami.task.agents:ResponseClassifierTeacher \
--model projects.light_whoami.agents.rpa_rerank:LongRPARerankAgent \
-mf zoo:light_whoami/vanilla_1024/model -ne 10
accuracy bleu-4 character_accuracy ctpb ctps exps exs f1 gpu_mem lr ltpb ltps tpb tps
0 1.231e-05 1 255.1 30.57 .1198 10 .1619 .2882 1e-09 25.6 3.067 280.7 33.63
Light Am I Me Or You Rpa Unlikelihood (128-Truncation) Model¶
RPA Unlikelihood-trained model from the Am I Me or You project.
Example invocation(s):
parlai em -mf zoo:light_whoami/rpa_ul_128/model \
-t light_dialog:simple_multi:light_label=speech \
-ne 10
accuracy bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem llen loss lr ltpb ltps ltrunc ltrunclen ppl right_class token_acc token_em 0 1.365e-05 255.1 126.7 8.552 .9000 128.4 .0675 10 .1767 .1945 25.6 2.376 7e-10 25.6 1.728 0 0 10.77 .7000 .4961 0
tpb tps ul_loss wrong_class wrong_class_all
152.3 10.28 .4625 .3000 .3000
Light Am I Me Or You Rpa Unlikelihood (1024-Truncation) Model¶
RPA Unlikelihood-trained model (1024-truncation) from the Am I Me or You project.
Example invocation(s):
parlai em -mf zoo:light_whoami/rpa_ul_1024/model \
-t light_dialog:simple_multi:light_label=speech \
-ne 10
accuracy bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem llen loss lr ltpb ltps ltrunc ltrunclen ppl right_class token_acc token_em 0 1.231e-05 255.1 255.1 14.99 0 0 .05876 10 .1897 .1947 25.6 2.327 7e-10 25.6 1.504 0 0 10.24 .9000 .5117 0
tpb tps ul_loss wrong_class wrong_class_all
280.7 16.49 .6992 .1000 .1000
Light Am I Me Or You Multi-Objective Model¶
Multi-Objective-trained model from the Am I Me or You project. This model can both generate a response, and identify the correct speaker given a conversational context and a candidate utterance
Example invocation(s):
parlai em -mf zoo:light_whoami/multiobjective/model \
-t projects.light_whoami.task.agents:MultiObjectiveTeacher \
-ne 10
accuracy bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 full_hits@1 full_mean_character_loss full_mrr full_rank gpu_mem llen loss lr ltpb ltps 0 1.231e-05 255.1 255.1 29.92 0 0 .1173 10 .1930 1 .04143 1 1 .2201 25.6 2.361 1e-09 25.6 3.003
ltrunc ltrunclen ppl token_acc token_em tpb tps
0 0 10.6 .5039 0 280.7 32.93
Light Am I Me Or You Profile Expanded Attention (128-Truncation) Model¶
A model that re-attends to select inputs from the context (specifically, character names and self persona)
Example invocation(s):
parlai em -mf zoo:light_whoami/profile_expanded_attention_128/model \
-t projects.light_whoami.task.agents:BaseSimpleMultiTeacher \
--mutators clean_context_mutator \
-ne 10
accuracy bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem llen loss lr ltpb ltps ltrunc ltrunclen ppl token_acc token_em tpb tps
0 .01958 254.5 126.5 8.154 .9000 128 .06446 10 .2210 .1580 25.6 2.296 7e-10 25.6 1.65 0 0 9.936 .5156 0 152.1 9.804
parlai em --predictor-model-file zoo:light_whoami/rpa_reranker/model \
-t projects.light_whoami.task.agents:ResponseClassifierTeacher \
--mutators clean_context_mutator \
--model projects.light_whoami.agents.expanded_attention:ExpandedDecoderAttentionAndRPARerankerAgent \
-mf zoo:light_whoami/profile_expanded_attention_128/model -ne 10
accuracy bleu-4 character_accuracy ctpb ctps exps exs f1 gpu_mem lr ltpb ltps tpb tps
0 .01958 1 126.5 7.976 .06305 10 .2198 .2309 7e-10 25.6 1.614 152.1 9.59
Light Am I Me Or You Profile Expanded Attention (1024-Truncation) Model¶
A model that re-attends to select inputs from the context (specifically, character names, self persona, and LIGHT setting)
Example invocation(s):
parlai em -mf zoo:light_whoami/profile_expanded_attention_1024/model \
-t projects.light_whoami.task.agents:BaseSimpleMultiTeacher \
--mutators clean_context_mutator \
-ne 10
accuracy bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem llen loss lr ltpb ltps ltrunc ltrunclen ppl token_acc token_em tpb tps
0 1.08e-05 254.5 254.5 17.32 0 0 .06804 10 .1514 .1583 25.6 2.285 7e-10 25.6 1.742 0 0 9.827 .5273 0 280.1 19.06
parlai em --predictor-model-file zoo:light_whoami/rpa_reranker/model \
-t projects.light_whoami.task.agents:ResponseClassifierTeacher \
--mutators clean_context_mutator \
--model projects.light_whoami.agents.expanded_attention:LongExpandedDecoderAttentionAndRPARerankerAgent \
-mf zoo:light_whoami/profile_expanded_attention_1024/model -ne 10
accuracy bleu-4 character_accuracy ctpb ctps exps exs f1 gpu_mem lr ltpb ltps tpb tps
0 1.08e-05 1 254.5 17.05 .06701 10 .1441 .2312 7e-10 25.6 1.715 280.1 18.77
Light Am I Me Or You Automated Expanded Attention (1024-Truncation) Model¶
A model that re-attends to select inputs from the context. An RPA classifier is used to select which parts of the input to re-attend to
Example invocation(s):
parlai em -mf zoo:light_whoami/automated_expanded_attention_1024/model \
-t projects.light_whoami.task.agents:BaseSimpleMultiTeacher \
--mutators share_self_character -ne 10
accuracy bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem llen loss lr ltpb ltps ltrunc ltrunclen ppl token_acc token_em tpb tps
0 1.066e-05 255.1 255.1 22.1 0 0 .08663 10 .1355 .1947 25.6 2.313 7e-10 25.6 2.218 0 0 10.1 .5195 0 280.7 24.32
Light Am I Me Or You Automated Expanded Attention + Multi-Objective Model¶
A model that re-attends to select inputs from the context. An RPA classifier is used to select which parts of the input to re-attend to
Example invocation(s):
parlai em -mf zoo:light_whoami/expanded_and_multiobjective_1024/model \
-t projects.light_whoami.task.agents:MultiObjectiveTeacher \
--mutators share_self_character,clean_context_mutator -ne 10
accuracy bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 full_hits@1 full_mean_character_loss full_mrr full_rank gpu_mem llen loss lr ltpb ltps 0 1.849e-08 254.5 254.5 22.72 0 0 .08928 10 .1628 .9000 2.245 .9002 43.6 .1589 25.6 2.347 7e-10 25.6 2.285
ltrunc ltrunclen ppl token_acc token_em tpb tps
0 0 10.45 .5078 0 280.1 25.01
Personality Captions models¶
Transresnet (Resnet 152) Personality-Captions Model¶
Transresnet Model pretrained on the Personality-Captions task
Example invocation(s):
parlai eval_model -t personality_captions -mf zoo:personality_captions/transresnet/model --num-test-labels 5 -dt test
{'exs': 10000, 'accuracy': 0.5113, 'f1': 0.5951, 'hits@1': 0.511, 'hits@5': 0.816, 'hits@10': 0.903, 'hits@100': 0.998, 'bleu': 0.4999, 'hits@1/100': 1.0, 'loss': -0.002, 'med_rank': 1.0}
Pretrained Transformers models¶
Poly-Encoder Transformer Reddit Pretrained Model¶
Poly-Encoder pretrained on Reddit. Use this model as an --init-model for a poly-encoder when fine-tuning on another task. For more details on how to train, see the project page.
Example invocation(s):
parlai train_model --init-model zoo:pretrained_transformers/poly_model_huge_reddit/model -t convai2 --model transformer/polyencoder --batchsize 256 --eval-batchsize 10 --warmup_updates 100 --lr-scheduler-patience 0 --lr-scheduler-decay 0.4 -lr 5e-05 --data-parallel True --history-size 20 --label-truncate 72 --text-truncate 360 --num-epochs 8.0 --max_train_time 200000 -veps 0.5 -vme 8000 --validation-metric accuracy --validation-metric-mode max --save-after-valid True --log_every_n_secs 20 --candidates batch --fp16 True --dict-tokenizer bpe --dict-lower True --optimizer adamax --output-scaling 0.06 --variant xlm --reduction-type mean --share-encoders False --learn-positional-embeddings True --n-layers 12 --n-heads 12 --ffn-size 3072 --attention-dropout 0.1 --relu-dropout 0.0 --dropout 0.1 --n-positions 1024 --embedding-size 768 --activation gelu --embeddings-scale False --n-segments 2 --learn-embeddings True --polyencoder-type codes --poly-n-codes 64 --poly-attention-type basic --dict-endtoken __start__ --model-file <YOUR MODEL FILE>
(subject to some variance, you may see the following as a result of validation of the model)
{'exs': 7801, 'accuracy': 0.8942 ...}
Poly-Encoder Transformer Wikipedia/Toronto Books Pretrained Model¶
Poly-Encoder pretrained on Wikipedia/Toronto Books. Use this model as an --init-model for a poly-encoder when fine-tuning on another task. For more details on how to train, see the project page.
Example invocation(s):
parlai train_model --init-model zoo:pretrained_transformers/poly_model_huge_wikito/model -t convai2 --model transformer/polyencoder --batchsize 256 --eval-batchsize 10 --warmup_updates 100 --lr-scheduler-patience 0 --lr-scheduler-decay 0.4 -lr 5e-05 --data-parallel True --history-size 20 --label-truncate 72 --text-truncate 360 --num-epochs 8.0 --max_train_time 200000 -veps 0.5 -vme 8000 --validation-metric accuracy --validation-metric-mode max --save-after-valid True --log_every_n_secs 20 --candidates batch --fp16 True --dict-tokenizer bpe --dict-lower True --optimizer adamax --output-scaling 0.06 --variant xlm --reduction-type mean --share-encoders False --learn-positional-embeddings True --n-layers 12 --n-heads 12 --ffn-size 3072 --attention-dropout 0.1 --relu-dropout 0.0 --dropout 0.1 --n-positions 1024 --embedding-size 768 --activation gelu --embeddings-scale False --n-segments 2 --learn-embeddings True --polyencoder-type codes --poly-n-codes 64 --poly-attention-type basic --dict-endtoken __start__ --model-file <YOUR MODEL FILE>
(subject to some variance, you may see the following as a result of validation of the model)
{'exs': 7801, 'accuracy': 0.861 ...}
Bi-Encoder Transformer Reddit Pretrained Model¶
Bi-Encoder pretrained on Reddit. Use this model as an --init-model for a bi-encoder when fine-tuning on another task. For more details on how to train, see the project page.
Example invocation(s):
parlai train_model --init-model zoo:pretrained_transformers/bi_model_huge_reddit/model --batchsize 512 -t convai2 --model transformer/biencoder --eval-batchsize 6 --warmup_updates 100 --lr-scheduler-patience 0 --lr-scheduler-decay 0.4 -lr 5e-05 --data-parallel True --history-size 20 --label-truncate 72 --text-truncate 360 --num-epochs 10.0 --max_train_time 200000 -veps 0.5 -vme 8000 --validation-metric accuracy --validation-metric-mode max --save-after-valid True --log_every_n_secs 20 --candidates batch --dict-tokenizer bpe --dict-lower True --optimizer adamax --output-scaling 0.06 --variant xlm --reduction-type mean --share-encoders False --learn-positional-embeddings True --n-layers 12 --n-heads 12 --ffn-size 3072 --attention-dropout 0.1 --relu-dropout 0.0 --dropout 0.1 --n-positions 1024 --embedding-size 768 --activation gelu --embeddings-scale False --n-segments 2 --learn-embeddings True --share-word-embeddings False --dict-endtoken __start__ --fp16 True --model-file <YOUR MODEL FILE>
(subject to some variance, you may see the following as a result of validation of the model)
{'exs': 7801, 'accuracy': 0.8686 ...}
Bi-Encoder Transformer Wikipedia/Toronto Books Pretrained Model¶
Bi-Encoder pretrained on Wikipedia/Toronto Books. Use this model as an --init-model for a poly-encoder when fine-tuning on another task. For more details on how to train, see the project page.
Example invocation(s):
parlai train_model --init-model zoo:pretrained_transformers/bi_model_huge_wikito/model --batchsize 512 -t convai2 --model transformer/biencoder --eval-batchsize 6 --warmup_updates 100 --lr-scheduler-patience 0 --lr-scheduler-decay 0.4 -lr 5e-05 --data-parallel True --history-size 20 --label-truncate 72 --text-truncate 360 --num-epochs 10.0 --max_train_time 200000 -veps 0.5 -vme 8000 --validation-metric accuracy --validation-metric-mode max --save-after-valid True --log_every_n_secs 20 --candidates batch --dict-tokenizer bpe --dict-lower True --optimizer adamax --output-scaling 0.06 --variant xlm --reduction-type mean --share-encoders False --learn-positional-embeddings True --n-layers 12 --n-heads 12 --ffn-size 3072 --attention-dropout 0.1 --relu-dropout 0.0 --dropout 0.1 --n-positions 1024 --embedding-size 768 --activation gelu --embeddings-scale False --n-segments 2 --learn-embeddings True --share-word-embeddings False --dict-endtoken __start__ --fp16 True --model-file <YOUR MODEL FILE>
(subject to some variance, you may see the following as a result of validation of the model)
{'exs': 7801, 'accuracy': 0.846 ...}
Cross-Encoder Transformer Reddit Pretrained Model¶
Cross-Encoder pretrained on Reddit. Use this model as an --init-model for a cross-encoder when fine-tuning on another task. For more details on how to train, see the project page.
Example invocation(s):
parlai train_model --init-model zoo:pretrained_transformers/cross_model_huge_reddit/model -t convai2 --model transformer/crossencoder --batchsize 16 --eval-batchsize 10 --warmup_updates 1000 --lr-scheduler-patience 0 --lr-scheduler-decay 0.4 -lr 5e-05 --data-parallel True --history-size 20 --label-truncate 72 --text-truncate 360 --num-epochs 12.0 --max_train_time 200000 -veps 0.5 -vme 2500 --validation-metric accuracy --validation-metric-mode max --save-after-valid True --log_every_n_secs 20 --candidates inline --fp16 True --dict-tokenizer bpe --dict-lower True --optimizer adamax --output-scaling 0.06 --variant xlm --reduction-type first --share-encoders False --learn-positional-embeddings True --n-layers 12 --n-heads 12 --ffn-size 3072 --attention-dropout 0.1 --relu-dropout 0.0 --dropout 0.1 --n-positions 1024 --embedding-size 768 --activation gelu --embeddings-scale False --n-segments 2 --learn-embeddings True --dict-endtoken __start__ --model-file <YOUR MODEL FILE>
(subject to some variance, you may see the following as a result of validation of the model)
{'exs': 7801, 'accuracy': 0.903 ...}
Cross-Encoder Transformer Wikipedia/Toronto Books Pretrained Model¶
Cross-Encoder pretrained on Wikipedia/Toronto Books. Use this model as an --init-model for a poly-encoder when fine-tuning on another task. For more details on how to train, see the project page.
Example invocation(s):
parlai train_model --init-model zoo:pretrained_transformers/cross_model_huge_wikito/model -t convai2 --model transformer/crossencoder --batchsize 16 --eval-batchsize 10 --warmup_updates 1000 --lr-scheduler-patience 0 --lr-scheduler-decay 0.4 -lr 5e-05 --data-parallel True --history-size 20 --label-truncate 72 --text-truncate 360 --num-epochs 12.0 --max_train_time 200000 -veps 0.5 -vme 2500 --validation-metric accuracy --validation-metric-mode max --save-after-valid True --log_every_n_secs 20 --candidates inline --fp16 True --dict-tokenizer bpe --dict-lower True --optimizer adamax --output-scaling 0.06 --variant xlm --reduction-type first --share-encoders False --learn-positional-embeddings True --n-layers 12 --n-heads 12 --ffn-size 3072 --attention-dropout 0.1 --relu-dropout 0.0 --dropout 0.1 --n-positions 1024 --embedding-size 768 --activation gelu --embeddings-scale False --n-segments 2 --learn-embeddings True --dict-endtoken __start__ --model-file <YOUR MODEL FILE>
(subject to some variance, you may see the following as a result of validation of the model)
{'exs': 7801, 'accuracy': 0.873 ...}
Convai2 models¶
Poly-Encoder Transformer Convai2 Model¶
Polyencoder pretrained on Reddit and fine-tuned on ConvAI2 scoring 89+ hits @ 1/20. See the pretrained_transformers directory for a list of other available pretrained transformers
Example invocation(s):
parlai interactive -mf zoo:pretrained_transformers/model_poly/model -t convai2
hi how are you doing ?
[Polyencoder]: i am alright . i am back from the library .
Enter Your Message: oh, what do you do for a living?
[Polyencoder]: i work at the museum downtown . i love it there .
Enter Your Message: what is your favorite drink?
[Polyencoder]: i am more of a tea guy . i get my tea from china .
parlai eval_model -mf zoo:pretrained_transformers/model_poly/model -t convai2 --eval-candidates inline
[ Finished evaluating tasks ['convai2'] using datatype valid ]
{'exs': 7801, 'accuracy': 0.8942, 'f1': 0.9065, 'hits@1': 0.894, 'hits@5': 0.99, 'hits@10': 0.997, 'hits@100': 1.0, 'bleu': 0.8941, 'lr': 5e-09, 'total_train_updates': 0, 'examples': 7801, 'loss': 3004.0, 'mean_loss': 0.385, 'mean_rank': 1.234, 'mrr': 0.9359}
Bi-Encoder Transformer Convai2 Model¶
Bi-encoder pretrained on Reddit and fine-tuned on ConvAI2 scoring ~87 hits @ 1/20.
Example invocation(s):
parlai interactive -mf zoo:pretrained_transformers/model_bi/model -t convai2
hi how are you doing ?
[Biencoder]: my mother is from russia .
Enter Your Message: oh cool, whereabouts ?
[Biencoder]: no , she passed away when i was 18 . thinking about russian recipes she taught me ,
Enter Your Message: what do you cook?
[Biencoder]: like meat mostly , me and my dogs love them , do you like dogs ?
parlai eval_model -mf zoo:pretrained_transformers/model_bi/model -t convai2 --eval-candidates inline
[ Finished evaluating tasks ['convai2'] using datatype valid ]
{'exs': 7801, 'accuracy': 0.8686, 'f1': 0.8833, 'hits@1': 0.869, 'hits@5': 0.987, 'hits@10': 0.996, 'hits@100': 1.0, 'bleu': 0.8685, 'lr': 5e-09, 'total_train_updates': 0, 'examples': 7801, 'loss': 28.77, 'mean_loss': 0.003688, 'mean_rank': 1.301, 'mrr': 0.9197}
Imageseq2Seq Dodecadialogue Convai2 Ft Model¶
Image Seq2Seq model trained on all DodecaDialogue tasks and fine-tuned on Convai2
Example invocation(s):
parlai interactive -mf zoo:dodecadialogue/convai2_ft/model -t convai2 --inference beam --beam-size 3 --beam-min-length 10 --beam-block-ngram 3 --beam-context-block-ngram 3
[context]: your persona: i currently work for ibm in chicago.
your persona: i'm not a basketball player though.
your persona: i am almost 7 feet tall.
your persona: i'd like to retire to hawaii in the next 10 years.
Enter Your Message: hi how's it going
[ImageSeq2seq]: i ' m doing well . how are you ?
Enter Your Message: i'm well, i am really tall
[ImageSeq2seq]: that ' s cool . i like simple jokes .
parlai eval_model -mf zoo:dodecadialogue/convai2_ft/model -t convai2
[ Finished evaluating tasks ['convai2'] using datatype valid ]
exs gpu_mem loss lr ppl token_acc total_train_updates tpb
7801 .2993 2.415 7.5e-06 11.19 .4741 15815 845.8
Unlikelihood Convai2 Context And Label Repetition Model¶
Dialogue model finetuned on ConvAI2 with context and label repetition unlikelihood
Example invocation(s):
python parlai/scripts/interactive.py -mf zoo:dialogue_unlikelihood/rep_convai2_ctxt_and_label/model -m projects.dialogue_unlikelihood.agents:RepetitionUnlikelihoodAgent
Enter Your Message: Hi.
[RepetitionUnlikelihood]: hi , how are you doing today ?
Unlikelihood Convai2 Context Repetition Model¶
Dialogue model finetuned on ConvAI2 with context repetition unlikelihood
Example invocation(s):
python parlai/scripts/interactive.py -mf zoo:dialogue_unlikelihood/rep_convai2_ctxt/model -m projects.dialogue_unlikelihood.agents:RepetitionUnlikelihoodAgent
Enter Your Message: Hi.
[RepetitionUnlikelihood]: hi , how are you doing today ?
Unlikelihood Convai2 Label Repetition Model¶
Dialogue model finetuned on ConvAI2 with label repetition unlikelihood
Example invocation(s):
python parlai/scripts/interactive.py -mf zoo:dialogue_unlikelihood/rep_convai2_label/model -m projects.dialogue_unlikelihood.agents:RepetitionUnlikelihoodAgent
Enter Your Message: Hi.
[RepetitionUnlikelihood]: hi , how are you doing today ?
Unlikelihood Vocab Alpha 1E0 Model¶
Dialogue model finetuned on convai2 with vocab unlikelihood, alpha value 1e0
Example invocation(s):
python parlai/scripts/interactive.py -mf zoo:dialogue_unlikelihood/vocab_alpha1e0/model -m projects.dialogue_unlikelihood.agents:TransformerSequenceVocabUnlikelihoodAgent
Enter Your Message: Hi.
[TransformerSequenceVocabUnlikelihood]: hi there ! how are you ?
Unlikelihood Vocab Alpha 1E1 Model¶
Dialogue model finetuned on convai2 with vocab unlikelihood, alpha value 1e1
Example invocation(s):
python parlai/scripts/interactive.py -mf zoo:dialogue_unlikelihood/vocab_alpha1e1/model -m projects.dialogue_unlikelihood.agents:TransformerSequenceVocabUnlikelihoodAgent
Enter Your Message: Hi.
[TransformerSequenceVocabUnlikelihood]: hi how are you today
Unlikelihood Vocab Alpha 1E2 Model¶
Dialogue model finetuned on convai2 with vocab unlikelihood, alpha value 1e2
Example invocation(s):
python parlai/scripts/interactive.py -mf zoo:dialogue_unlikelihood/vocab_alpha1e2/model -m projects.dialogue_unlikelihood.agents:TransformerSequenceVocabUnlikelihoodAgent
Enter Your Message: Hi.
[TransformerSequenceVocabUnlikelihood]: hello , how are you ?
Unlikelihood Vocab Alpha 1E3 Model¶
Dialogue model finetuned on convai2 with vocab unlikelihood, alpha value 1e3
Example invocation(s):
python parlai/scripts/interactive.py -mf zoo:dialogue_unlikelihood/vocab_alpha1e3/model -m projects.dialogue_unlikelihood.agents:TransformerSequenceVocabUnlikelihoodAgent
Enter Your Message: Hi.
[TransformerSequenceVocabUnlikelihood]: hi there !
Image Chat models¶
Transresnet (Resnet152) Image-Chat Model¶
Transresnet Multimodal Model pretrained on the Image-Chat task
Example invocation(s):
parlai eval_model -t image_chat -mf zoo:image_chat/transresnet_multimodal/model -dt test
{'exs': 29991, 'accuracy': 0.4032, 'f1': 0.4432, 'hits@1': 0.403, 'hits@5': 0.672, 'hits@10': 0.779, 'hits@100': 1.0, 'bleu': 0.3923,'first_round': {'hits@1/100': 0.3392, 'loss': -0.002001, 'med_rank': 3.0},'second_round': {'hits@1/100': 0.4558, 'loss': -0.002001, 'med_rank': 2.0},'third_round+': {'hits@1/100': 0.4147, 'loss': -0.002001, 'med_rank': 2.0}}'hits@10': 0.903, 'hits@100': 0.998, 'bleu': 0.4999, 'hits@1/100': 1.0, 'loss': -0.002, 'med_rank': 1.0}
Imageseq2Seq Dodecadialogue Image Chat Ft Model¶
Image Seq2Seq model trained on all DodecaDialogue tasks and fine-tuned on the Image Chat task
Example invocation(s):
parlai eval_model -mf zoo:dodecadialogue/image_chat_ft/model -t image_chat:generation --image-mode no_image_model
[ Finished evaluating tasks ['image_chat:generation'] using datatype valid ]
exs gpu_mem loss lr ppl token_acc total_train_updates tpb
15000 .2231 4.353 3.125e-07 77.73 .2905 321001 1653
Dialogue Safety models¶
Transformer Classifier Single-Turn Dialogue Safety Model¶
Classifier trained on both the standard and adversarial safety tasks in addition to Wikipedia Toxic Comments.
Example invocation(s):
parlai eval_model -t dialogue_safety:adversarial --round 3 -dt test -mf zoo:dialogue_safety/single_turn/model -bs 40
{'exs': 3000, 'accuracy': 0.9627, 'f1': 0.9627, 'bleu': 9.627e-10, 'lr': 5e-09, 'total_train_updates': 0, 'examples': 3000, 'mean_loss': 0.005441, 'class___notok___recall': 0.7833, 'class___notok___prec': 0.8333, 'class___notok___f1': 0.8076, 'class___ok___recall': 0.9826, 'class___ok___prec': 0.9761, 'class___ok___f1': 0.9793, 'weighted_f1': 0.9621}
Bert Classifier Multi-Turn Dialogue Safety Model¶
Classifier trained on the multi-turn adversarial safety task in addition to both the single-turn standard and adversarial safety tasks and Wikipedia Toxic Comments.
Example invocation(s):
parlai eval_model -t dialogue_safety:multiturn -dt test -mf zoo:dialogue_safety/multi_turn/model --split-lines True -bs 40
{'exs': 3000, 'accuracy': 0.9317, 'f1': 0.9317, 'bleu': 9.317e-10, 'lr': 5e-09, 'total_train_updates': 0, 'examples': 3000, 'mean_loss': 0.008921, 'class___notok___recall': 0.7067, 'class___notok___prec': 0.6444, 'class___notok___f1': 0.6741, 'class___ok___recall': 0.9567, 'class___ok___prec': 0.9671, 'class___ok___f1': 0.9618, 'weighted_f1': 0.9331}
Integration Tests models¶
Integration Test Models¶
Model files used to check backwards compatibility and code coverage of important standard models.
Example invocation(s):
parlai eval_model -mf zoo:unittest/transformer_generator2/model -t integration_tests:multiturn_candidate -m transformer/generator
{'exs': 400, 'accuracy': 1.0, 'f1': 1.0, 'bleu-4': 0.2503, 'lr': 0.001, 'total_train_updates': 5000, 'gpu_mem_percent': 9.37e-05, 'loss': 0.0262, 'token_acc': 1.0, 'nll_loss': 7.935e-05, 'ppl': 1.0}
#Dodeca models¶
Imageseq2Seq Dodecadialogue All Tasks Mt Model¶
Image Seq2Seq model trained on all DodecaDialogue tasks
Example invocation(s):
parlai interactive -mf zoo:dodecadialogue/all_tasks_mt/model --inference beam --beam-size 3 --beam-min-length 10 --beam-block-ngram 3 --beam-context-block-ngram 3
Enter Your Message: hi how are you?
[ImageSeq2seq]: i ' m doing well . how are you ?
Enter Your Message: not much, what do you like to do?
[ImageSeq2seq]: i like to go to the park and play with my friends .
parlai eval_model -mf zoo:dodecadialogue/all_tasks_mt/model -t "#Dodeca"--prepend-personality True --prepend-gold-knowledge True --image-mode no_image_model
[ Finished evaluating tasks ['#Dodeca'] using datatype valid ]
exs gpu_mem loss lr ppl token_acc total_train_updates tpb
WizTeacher 3939 2.161 8.678 .5325
all 91526 .3371 2.807 9.375e-07 18.23 .4352 470274 2237
convai2 7801 2.421 11.26 .4721
cornell_movie 13905 3.088 21.93 .4172
dailydialog 8069 2.47 11.82 .4745
empathetic_dialogues 5738 2.414 11.18 .4505
igc 486 2.619 13.73 .4718
image_chat:Generation 15000 3.195 24.42 .3724
light_dialog 6623 2.944 19 .3918
twitter 10405 3.61 36.98 .3656
ubuntu 19560 3.148 23.3 .4035
Imageseq2Seq Dodecadialogue Base Model¶
Image Seq2Seq base model, from which all DodecaDialogue models were trained
Example invocation(s):
parlai train_model -t "#Dodeca" --prepend-gold-knowledge true --prepend-personality true -mf /tmp/dodeca_model --init-model zoo:dodecadialogue/base_model/model --dict-file zoo:dodecadialogue/dict/dodeca.dict --model image_seq2seq --dict-tokenizer bpe --dict-lower true -bs 32 -eps 0.5 -esz 512 --ffn-size 2048 --fp16 false --n-heads 16 --n-layers 8 --n-positions 512 --text-truncate 512 --label-truncate 128 --variant xlm -lr 7e-6 --lr-scheduler reduceonplateau --optimizer adamax --dropout 0.1 --validation-every-n-secs 3600 --validation-metric ppl --validation-metric-mode min --validation-patience 10 --activation gelu --embeddings-scale true --learn-positional-embeddings true --betas 0.9,0.999 --warmup-updates 2000 --gradient-clip 0.1
A trained model (logs omitted)
Cornell Movie models¶
Imageseq2Seq Dodecadialogue Cornell Movie Ft Model¶
Image Seq2Seq model trained on all DodecaDialogue tasks and fine-tuned on the Cornell Movie task
Example invocation(s):
parlai interactive -mf zoo:dodecadialogue/cornell_movie_ft/model --inference beam --beam-size 10 --beam-min-length 20 --beam-block-ngram 3 --beam-context-block-ngram 3
Enter Your Message: hi how's it going?
[ImageSeq2seq]: oh , it ' s great . i ' m having a great time . how are you doing ?
Enter Your Message: i'm doing well, what do you like to do?
[ImageSeq2seq]: i like to go to the movies . what about you ? do you have any hobbies ?
parlai eval_model -mf zoo:dodecadialogue/cornell_movie_ft/model -t cornell_movie
[ Finished evaluating tasks ['cornell_movie'] using datatype valid ]
exs gpu_mem loss lr ppl token_acc total_train_updates tpb
13905 .07094 2.967 2.5e-06 19.43 .4290 29496 15.76
Dailydialog models¶
Imageseq2Seq Dodecadialogue Dailydialog Ft Model¶
Image Seq2Seq model trained on all DodecaDialogue tasks and fine-tuned on the DailyDialog task
Example invocation(s):
parlai interactive -mf zoo:dodecadialogue/daily_dialog_ft/model --inference beam --beam-size 5 --beam-min-length 10 --beam-block-ngram 3 --beam-context-block-ngram 3
Enter Your Message: hi how's it going
[ImageSeq2seq]: i ' m doing well . how about you ?
Enter Your Message: not much, what do you like to do?
[ImageSeq2seq]: i like to go to the beach and play volleyball .
parlai eval_model -mf zoo:dodecadialogue/daily_dialog_ft/model -t dailydialog
[ Finished evaluating tasks ['dailydialog'] using datatype valid ]
exs gpu_mem loss lr ppl token_acc total_train_updates tpb
8069 .06787 2.326 7.5e-06 10.24 .5093 150959 15.67
TBD models¶
Imageseq2Seq Dodecadialogue Eli5 Ft Model¶
Image Seq2Seq model trained on all DodecaDialogue tasks and fine-tuned on the ELI5 task. The ELI5 model in the zoo is the one that scored the highest ROUGE scores; it may have slightly lower f1, bleu, and slightly higher ppl than the numbers reported in the dodeca paper.
Example invocation(s):
parlai interactive -mf zoo:dodecadialogue/eli5_ft/model --inference beam --beam-size 10 --beam-min-length 200 --beam-block-ngram 3 --beam-context-block-ngram 3
Enter Your Message: Hi, can you tell me about quantum physics?
[ImageSeq2seq]: yes , i can . quantum physics is the study of how particles interact with each other , and how they interact with other particles . it ' s important to note that quantum mechanics is n ' t the same as classical physics . classical physics is a study of the properties of particles , and what they do . in classical physics , there are two types of particles : quarks and neutrinos . quarks are made up of quarks , neutrinos , and electrons . neutrinos are made of protons , neutrons , electrons , and neutrons . they ' re all the same thing , but they all have the same properties . so , if you ' re interested in quantum physics , you might want to check out / r / askscience . there ' s a subreddit for that sort of thing . edit : i ' m not sure what you mean by " quantum physics " , but i ' ll let you know if you want to know more . edit 2 : thanks for the gold !
Imageseq2Seq Dodecadialogue Pushshift.Io Reddit Ft Model¶
Image Seq2Seq model trained on all DodecaDialogue tasks and fine-tuned on the pushshift.io Reddit task
Example invocation(s):
parlai interactive -mf zoo:dodecadialogue/reddit_ft/model --inference beam --beam-size 5 --beam-min-length 20 --beam-block-ngram 3 --beam-context-block-ngram 3
Enter Your Message: hi how's it going?
[ImageSeq2seq]: hi , i ' m doing pretty well . how are you ? : ) and yourself ? : d
Enter Your Message: just hanging in there, you up to anything fun?
[ImageSeq2seq]: not really . i just got home from work . i ' ll be back in a few hours .
Generative Pre-Trained Transformer 3¶
An API wrapper around OpenAI’s GPT-3 Model.
Example invocation(s):
parlai interactive -m gpt3 --openai-api-key <your_openai_api_key> --max-tokens 40 --model-name text-davinci-003
[context]: Hi!
[Gpt3Agent_1]: Hello human!
[Gpt3Agent_2]: Hello human!
Empathetic Dialogues models¶
Imageseq2Seq Dodecadialogue Empathetic Dialogue Ft Model¶
Image Seq2Seq model trained on all DodecaDialogue tasks and fine-tuned on the Empathetic Dialogue task
Example invocation(s):
parlai interactive -mf zoo:dodecadialogue/empathetic_dialogues_ft/model --inference beam --beam-size 5 --beam-min-length 10 --beam-block-ngram 3 --beam-context-block-ngram 3
Enter Your Message: hi, how's it going?
[ImageSeq2seq]: i ' m doing well . how are you ?
Enter Your Message: i'm fine, feeling a little sad
[ImageSeq2seq]: that ' s too bad . what ' s going on ?
parlai eval_model -mf zoo:dodecadialogue/empathetic_dialogues_ft/model -t empathetic_dialogues
[ Finished evaluating tasks ['empathetic_dialogues'] using datatype valid ]
exs gpu_mem loss lr ppl token_acc total_train_updates tpb
5738 .3278 2.405 7.5e-06 11.08 .4517 20107 1914
Igc models¶
Imageseq2Seq Dodecadialogue Image Grounded Conversations Ft Model¶
Image Seq2Seq model trained on all DodecaDialogue tasks and fine-tuned on the Image Grounded Conversations task
Example invocation(s):
parlai eval_model -mf zoo:dodecadialogue/igc_ft/model -t igc:responseOnly
[ Finished evaluating tasks ['igc:responseOnly'] using datatype valid ]
exs gpu_mem loss lr ppl token_acc total_train_updates tpb
162 .0726 2.832 1e-06 16.98 .4405 10215 9.852
Twitter models¶
Imageseq2Seq Dodecadialogue Twitter Ft Model¶
Image Seq2Seq model trained on all DodecaDialogue tasks and fine-tuned on the Twitter task
Example invocation(s):
parlai interactive -mf zoo:dodecadialogue/twitter_ft/model --inference beam --beam-size 10 --beam-min-length 20 --beam-block-ngram 3 --beam-context-block-ngram 3
Enter Your Message: hi how's it going?
[ImageSeq2seq]: it ' s going well ! how are you ? @ smiling_face_with_heart - eyes @
Enter Your Message: im doing well, what do you like to do
[ImageSeq2seq]: hi ! i ' m doing well ! i like to read , watch movies , play video games , and listen to music . how about you ?
parlai eval_model -mf zoo:dodecadialogue/twitter_ft/model -t twitter
[ Finished evaluating tasks ['twitter'] using datatype valid ]
exs gpu_mem loss lr ppl token_acc total_train_updates tpb
10405 .3807 3.396 7.5e-06 29.83 .3883 524029 2395
Ubuntu models¶
Imageseq2Seq Dodecadialogue Ubuntu V2 Ft Model¶
Image Seq2Seq model trained on all DodecaDialogue tasks and fine-tuned on the Ubuntu V2 task
Example invocation(s):
parlai interactive -mf zoo:dodecadialogue/ubuntu_ft/model --inference beam --beam-size 2 --beam-min-length 10 --beam-block-ngram 3 --beam-context-block-ngram 3
Enter Your Message: hi how's it going?
[ImageSeq2seq]: i ' m fine . . . you ? .
Enter Your Message: doing ok, what do you like to do?
[ImageSeq2seq]: i like to read , write , and read .
parlai eval_model -mf zoo:dodecadialogue/ubuntu_ft/model -t ubuntu
[ Finished evaluating tasks ['ubuntu'] using datatype valid ]
exs gpu_mem loss lr ppl token_acc total_train_updates tpb
19560 .3833 2.844 2.5e-05 17.18 .4389 188076 3130
Blended Skill Talk models¶
Blendedskilltalk: Blendedskilltalk Single-Task Model¶
Pretrained polyencoder retrieval model fine-tuned on the BlendedSkillTalk dialogue task.
Example invocation(s):
parlai interactive -mf zoo:blended_skill_talk/bst_single_task/model -t blended_skill_talk
Results vary.
parlai eval_model -mf zoo:blended_skill_talk/bst_single_task/model -t blended_skill_talk -dt test
09:51:57 | Finished evaluating tasks ['blended_skill_talk'] using datatype test
accuracy bleu-4 exs f1 gpu_mem hits@1 hits@10 hits@100 hits@5 loss mrr rank tpb
.7920 .7785 5482 .8124 .0370 .7920 .9788 1 .9542 .8251 .8636 1.866 19.76
Blendedskilltalk: Convai2 Single-Task Model¶
Pretrained polyencoder retrieval model fine-tuned on the ConvAI2 dialogue task.
Example invocation(s):
parlai eval_model -mf zoo:blended_skill_talk/convai2_single_task/model -t blended_skill_talk -dt test
10:23:53 | Finished evaluating tasks ['blended_skill_talk'] using datatype test
accuracy bleu-4 exs f1 gpu_mem hits@1 hits@10 hits@100 hits@5 loss mrr rank tpb
.7678 .7553 5482 .7902 .07928 .7678 .9728 1 .9414 .9337 .8451 2.04 19.76
Blendedskilltalk: Empatheticdialogues Single-Task Model¶
Pretrained polyencoder retrieval model fine-tuned on the EmpatheticDialogues dialogue task.
Example invocation(s):
parlai eval_model -mf zoo:blended_skill_talk/ed_single_task/model -t blended_skill_talk -dt test
10:16:47 | Finished evaluating tasks ['blended_skill_talk'] using datatype test
accuracy bleu-4 exs f1 gpu_mem hits@1 hits@10 hits@100 hits@5 loss mrr rank tpb
.6895 .6774 5482 .7219 .07928 .6895 .9509 1 .9051 1.242 .7849 2.79 19.76
Blendedskilltalk: Wizard Of Wikipedia Single-Task Model¶
Pretrained polyencoder retrieval model fine-tuned on the Wizard of Wikipedia dialogue task.
Example invocation(s):
parlai eval_model -mf zoo:blended_skill_talk/wizard_single_task/model -t blended_skill_talk -dt test
10:34:46 | Finished evaluating tasks ['blended_skill_talk'] using datatype test
accuracy bleu-4 exs f1 gpu_mem hits@1 hits@10 hits@100 hits@5 loss mrr rank tpb
.6742 .6616 5482 .7059 .07928 .6742 .9445 1 .8902 1.321 .7706 2.962 19.76
Blendedskilltalk: Mt Single-Skills Model¶
Pretrained polyencoder retrieval model fine-tuned on the ConvAI2, EmpatheticDialogues, and Wizard of Wikipedia dialogue tasks.
Example invocation(s):
parlai eval_model -mf zoo:blended_skill_talk/multi_task/model -t blended_skill_talk -dt test
10:23:35 | Finished evaluating tasks ['blended_skill_talk'] using datatype test
accuracy bleu-4 exs f1 gpu_mem hits@1 hits@10 hits@100 hits@5 loss mrr rank tpb
.8010 .7872 5482 .8204 .07928 .8010 .9779 1 .9564 .8154 .8697 1.908 19.76
Blendedskilltalk: Mt Single-Skills Model Fine-Tuned On Bst¶
Pretrained polyencoder retrieval model fine-tuned on the ConvAI2, EmpatheticDialogues, and Wizard of Wikipedia dialogue tasks, and then further fine-tuned on the BlendedSkillTalk dialogue task.
Example invocation(s):
parlai eval_model -mf zoo:blended_skill_talk/multi_task_bst_tuned/model -t blended_skill_talk -dt test
10:36:01 | Finished evaluating tasks ['blended_skill_talk'] using datatype test
accuracy bleu-4 exs f1 gpu_mem hits@1 hits@10 hits@100 hits@5 loss mrr rank tpb
.8378 .8230 5482 .8543 .07928 .8378 .9872 1 .9704 .5897 .8963 1.604 19.76
Blender 90M¶
90< parameter generative model finetuned on blended_skill_talk tasks.
Example invocation(s):
python parlai/scripts/safe_interactive.py -mf zoo:blender/blender_90M/model -t blended_skill_talk
Enter Your Message: Hi what's up?
[TransformerGenerator]: hello , how are you ? i just got back from working at a law firm , how about you ?
Blender 2.7B¶
2.7B parameter generative model finetuned on blended_skill_talk tasks.
Example invocation(s):
python parlai/scripts/safe_interactive.py -mf zoo:blender/blender_3B/model -t blended_skill_talk
Enter Your Message: Hi how are you?
[TransformerGenerator]: I'm doing well. How are you doing? What do you like to do in your spare time?
Blender 1B Distilled¶
2.7B parameter generative model finetuned on blended_skill_talk tasks and then distilled to ~1.4B parameters and roughly 2x the inference speed.
Example invocation(s):
python parlai/scripts/safe_interactive.py -mf zoo:blender/blender_1Bdistill/model -t blended_skill_talk
Enter Your Message: Hi how are you?
[TransformerGenerator]: I'm doing well. How about yourself? What do you do for a living? I'm a creative writer.
Blender 400M Distilled¶
2.7B parameter generative model finetuned on blended_skill_talk tasks and then distilled to ~360M parameters and roughly 5x the inference speed.
Example invocation(s):
python parlai/scripts/safe_interactive.py -mf zoo:blender/blender_400Mdistill/model -t blended_skill_talk
Enter Your Message: Hi how are you?
[TransformerGenerator]: I'm doing well. How about you? What do you like to do in your free time?
Blender 9.4B¶
9.4B parameter generative model finetuned on blended_skill_talk tasks.
Example invocation(s):
python parlai/scripts/safe_interactive.py -mf zoo:blender/blender_9B/model -t blended_skill_talk
Enter Your Message: Hi!
[TransformerGenerator]: What do you do for a living? I'm a student at Miami University.
Multi-Modal Blenderbot (Mmb Degenpos)¶
Model trained to talk about both images and general chitchat, trained with a degendering teacher and with 75% of Image-Chat styles replaced by a generic polarity string
Example invocation(s):
python parlai/scripts/safe_interactive.py -t blended_skill_talk -mf ${FINETUNED_MODEL_PATH} --model projects.multimodal_blenderbot.agents:BiasAgent --delimiter $'
' --beam-block-ngram 3 --beam-context-block-ngram 3 --beam-min-length 20 --beam-size 10 --inference beam --model-parallel False
(results will vary)
Blenderbot3B With Name-Scrambling Gender-Bias Reduction¶
BlenderBot3B model tuned with the name-scrambling technique to have lower gender bias.
Example invocation(s):
parlai display_model -mf zoo:dialogue_bias/gender__name_scrambling/model -t blended_skill_talk --beam-block-full-context True --interactive-mode True
your persona: i work as an electrician.
your persona: i always sleep 8 hours a day.
Electrician
That sounds dangerous. Is it worth doing such a dangerous job?
Wekk it is okay is you are well trained. There are three levels: Apprentice, journeyman and Master.
Which level are you at?
labels: I received on-the-job training when i first started
model: I am a journeyman. I have been working in the trade for a long time.
Blenderbot3B With Token-Bin Control-Generation Gender-Bias Reduction¶
BlenderBot3B model tuned with the control-generation technique (based on the amount of gender over-indexing per token bin) to have lower gender bias.
Example invocation(s):
parlai display_model -mf zoo:dialogue_bias/gender__ctrl_gen_tokens/model --model projects.dialogue_bias.agents:NoBiasStyleGenAgent -t blended_skill_talk --beam-block-full-context True --interactive-mode True
your persona: i work as an electrician.
your persona: i always sleep 8 hours a day.
Electrician
That sounds dangerous. Is it worth doing such a dangerous job?
Wekk it is okay is you are well trained. There are three levels: Apprentice, journeyman and Master.
Which level are you at?
labels: I received on-the-job training when i first started
model: I am a Master Electrician. I work in a power plant. How about you?
Blenderbot3B With Sequence-Level Unlikelihood-Training Gender-Bias Reduction¶
BlenderBot3B model tuned with the (sequence-level) unlikelihood training technique to have lower gender bias.
Example invocation(s):
parlai display_model -mf zoo:dialogue_bias/gender__unlikelihood_sequence_level/model --model transformer/generator -t blended_skill_talk --beam-block-full-context True --interactive-mode True
your persona: i work as an electrician.
your persona: i always sleep 8 hours a day.
Electrician
That sounds dangerous. Is it worth doing such a dangerous job?
Wekk it is okay is you are well trained. There are three levels: Apprentice, journeyman and Master.
Which level are you at?
labels: I received on-the-job training when i first started
model: I'm a journeyman electrician. I work in the construction industry. How about you?
Blenderbot3B With Name-Scrambling Gender/Ethnicity-Bias Reduction¶
BlenderBot3B model tuned with the name-scrambling technique to have lower gender and race/ethnicity bias.
Example invocation(s):
parlai display_model -mf zoo:dialogue_bias/gender_ethnicity__name_scrambling/model -t blended_skill_talk --beam-block-full-context True --interactive-mode True
your persona: i work as an electrician.
your persona: i always sleep 8 hours a day.
Electrician
That sounds dangerous. Is it worth doing such a dangerous job?
Wekk it is okay is you are well trained. There are three levels: Apprentice, journeyman and Master.
Which level are you at?
labels: I received on-the-job training when i first started
model: I am a journeyman electrician. I work in the construction industry. How about you?
Pushshift.Io models¶
Tutorial Transformer Generator¶
Small (87M parameter) generative transformer, pretrained on pushshift.io Reddit.
Example invocation(s):
parlai interactive -mf zoo:tutorial_transformer_generator/model
Enter Your Message: hi, how are you today?
[TransformerGenerator]: i ' m doing well , how about you ?
Enter Your Message: I'm giving a tutorial on chatbots!
[TransformerGenerator]: that ' s awesome ! what ' s it about ?
Enter Your Message: bots just like you
[TransformerGenerator]: i ' ll be sure to check it out !
Reddit 2.7B¶
2.7B parameter generative model pretrained on Reddit but not finetuned.
Example invocation(s):
parlai train_model -t blended_skill_talk,wizard_of_wikipedia,convai2:normalized,empathetic_dialogues --multitask-weights 1,3,3,3 -veps 0.25 --attention-dropout 0.0 --batchsize 128 --model transformer/generator --embedding-size 2560 --ffn-size 10240 --variant prelayernorm --n-heads 32 --n-positions 128 --n-encoder-layers 2 --n-decoder-layers 24 --history-add-global-end-token end --delimiter ' ' --dict-tokenizer bytelevelbpe --dropout 0.1 --fp16 True --init-model zoo:blender/reddit_3B/model --dict-file zoo:blender/reddit_3B/model.dict --label-truncate 128 --log_every_n_secs 10 -lr 7e-06 --lr-scheduler reduceonplateau --lr-scheduler-patience 3 --optimizer adam --relu-dropout 0.0 --activation gelu --model-parallel true --save-after-valid True --text-truncate 128 --truncate 128 --warmup_updates 100 --fp16-impl mem_efficient --update-freq 2 --gradient-clip 0.1 --skip-generation True -vp 10 -vmt ppl -vmm min --model-file /tmp/test_train_27B
Results vary.
Reddit 9.4B¶
9.4B parameter generative model pretrained on Reddit but not finetuned.
Example invocation(s):
parlai train_model -t blended_skill_talk,wizard_of_wikipedia,convai2:normalized,empathetic_dialogues --multitask-weights 1,3,3,3 -veps 0.25 --attention-dropout 0.0 --batchsize 8 --eval-batchsize 64 --model transformer/generator --embedding-size 4096 --ffn-size 16384 --variant prelayernorm --n-heads 32 --n-positions 128 --n-encoder-layers 4 --n-decoder-layers 32 --history-add-global-end-token end --dict-tokenizer bytelevelbpe --dropout 0.1 --fp16 True --init-model zoo:blender/reddit_9B/model --dict-file zoo:blender/reddit_9B/model.dict --label-truncate 128 -lr 3e-06 -dynb full --lr-scheduler cosine --max-lr-steps 9000 --lr-scheduler-patience 3 --optimizer adam --relu-dropout 0.0 --activation gelu --model-parallel true --save-after-valid False --text-truncate 128 --truncate 128 --warmup_updates 1000 --fp16-impl mem_efficient --update-freq 4 --log-every-n-secs 30 --gradient-clip 0.1 --skip-generation True -vp 10 --max-train-time 84600 -vmt ppl -vmm min --model-file /tmp/test_train_94B
Results vary.
Wikipedia Plus Toronto Books models¶
Bart¶
BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension.
Example invocation(s):
parlai eval_model -mf zoo:bart/bart_large/model -t convai2 -bs 64
Finished evaluating tasks ['convai2'] using datatype valid
accuracy bleu-4 exs f1 gpu_mem loss ppl token_acc tpb
0 .0004641 7801 .02084 .4878 5.041 154.6 .2042 1652
parlai train_model -m bart -mf /tmp/model_file -t convai2 -bs 24 --fp16 true -eps 1 -lr 1e-5 --optimizer adam
valid:
accuracy bleu-4 exs f1 gpu_mem loss lr ppl token_acc total_train_updates tpb
.0001282 .01229 7801 .2035 .6361 2.386 1e-05 10.87 .4741 5478 321.3
Eli5 models¶
Unlikelihood Eli5 Context And Label Repetition Model¶
Dialogue model finetuned on ELI5 with context and label repetition unlikelihood
Example invocation(s):
python parlai/scripts/interactive.py -mf zoo:dialogue_unlikelihood/rep_eli5_ctxt_and_label/model -m projects.dialogue_unlikelihood.agents:RepetitionUnlikelihoodAgent
Enter Your Message: Hi.
[RepetitionUnlikelihood]: hi .
Unlikelihood Eli5 Context Repetition Model¶
Dialogue model finetuned on ELI5 with context repetition unlikelihood
Example invocation(s):
python parlai/scripts/interactive.py -mf zoo:dialogue_unlikelihood/rep_eli5_ctxt/model -m projects.dialogue_unlikelihood.agents:RepetitionUnlikelihoodAgent
Enter Your Message: Hi.
[RepetitionUnlikelihood]: hi .
Unlikelihood Eli5 Label Repetition Model¶
Dialogue model finetuned on ELI5 with label repetition unlikelihood
Example invocation(s):
python parlai/scripts/interactive.py -mf zoo:dialogue_unlikelihood/rep_eli5_label/model -m projects.dialogue_unlikelihood.agents:RepetitionUnlikelihoodAgent
Enter Your Message: Hi.
[RepetitionUnlikelihood]: hi .
Style Gen models¶
Style-Controlled Generation: C75-D+ Generator¶
Generator trained on dialogue datasets, with 75% of train examples appended with Image-Chat personality labels
Example invocation(s):
parlai eval_model --datatype test --model projects.style_gen.style_gen:StyleGenAgent --model-file zoo:style_gen/c75_labeled_dialogue_generator/model --skip-generation True --task style_gen:LabeledBlendedSkillTalk --use-style-frac 1.00
16:56:52 | Finished evaluating tasks ['style_gen:LabeledBlendedSkillTalk'] using datatype test
ctpb ctps exps exs gpu_mem loss ltpb ltps ppl token_acc tpb tps
120 1855 15.46 5482 .1635 2.248 19.94 308.2 9.468 .4872 139.9 2163
Style-Controlled Generation: Previous And Current Utterance Classifier¶
Classifier trained on Image-Chat turns 2 and 3 to classify the personality of an example given that utterance and the previous utterance.
Example invocation(s):
parlai eval_model --task style_gen:PrevCurrUttStyle --wrapper-task style_gen:LabeledBlendedSkillTalk --model-file zoo:style_gen/prev_curr_classifier/model --model projects.style_gen.classifier:ClassifierAgent --classes-from-file image_chat_personalities_file
18:42:33 | Finished evaluating tasks ['style_gen:PrevCurrUttStyle'] using datatype valid
accuracy bleu-4 ctpb ctps exps exs f1 gpu_mem loss lr ltpb ltps tpb tps
.9973 .01745 38.08 604.1 15.86 5651 .9973 .1622 2.129 5e-10 5.633 89.36 43.71 693.4
Style-Controlled Generation: Current-Utterance-Only Classifier¶
Classifier trained on Image-Chat turns 2 and 3 to classify the personality of an example given that utterance as the sole context.
Example invocation(s):
parlai eval_model --task style_gen:CurrUttOnlyStyle --wrapper-task style_gen:LabeledBlendedSkillTalk --model-file zoo:style_gen/curr_only_classifier/model --model projects.style_gen.classifier:ClassifierAgent --classes-from-file image_chat_personalities_file
16:46:41 | Finished evaluating tasks ['style_gen:CurrUttOnlyStyle'] using datatype valid
accuracy bleu-4 <PER_CLASS_METRICS_SNIPPED> clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem llen loss lr ltpb ltps ltrunc ltrunclen tpb tps .4311 .004642 19.18 19.18 425.9 0 0 22.2 5651 .4363 .1586 5.633 2.658 5e-10 5.633 125.1 0 0 24.82 550.9
weighted_f1
.4319
Bot Adversarial Dialogue models¶
Transformer Classifier Multi-Turn Dialogue Safety Model¶
Classifier trained on the filtered multi-turn bot adversarial dialogues in addition to both dialogue_safety single-turn standard and adversarial safety tasks and Wikipedia Toxic Comments.
Example invocation(s):
parlai eval_model -t bot_adversarial_dialogue:bad_num_turns=4 -dt test -mf zoo:bot_adversarial_dialogue/multi_turn_v0/model -bs 128
{'exs': 2598, 'accuracy': 0.8414, 'f1': 0.8414, 'loss': 0.5153, 'bleu-4': 8.414e-10, 'class___notok___recall': 0.8093, 'class___notok___prec': 0.7671, 'class___notok___f1': 0.7876, 'class___ok___recall': 0.8597, 'class___ok___prec': 0.8876, 'class___ok___f1': 0.8735, 'weighted_f1': 0.8423}
Transformer Classifier Multi-Turn Dialogue Safety Model¶
Classifier trained on the truncated multi-turn bot adversarial dialogues in addition to both dialogue_safety single-turn standard and adversarial safety tasks and Wikipedia Toxic Comments.
Example invocation(s):
parlai eval_model -t bot_adversarial_dialogue:bad_num_turns=4 -dt test -mf zoo:bot_adversarial_dialogue/multi_turn/model -bs 128
{'exs': 2598, 'accuracy': 0.8507, 'f1': 0.8507, 'loss': 0.3878, 'bleu-4': 8.507e-10, 'class___notok___recall': 0.8633, 'class___notok___prec': 0.7588, 'class___notok___f1': 0.8077, 'class___ok___recall': 0.8434, 'class___ok___prec': 0.9154, 'class___ok___f1': 0.8779, 'weighted_f1': 0.8524}
Sensitive Topics Evaluation models¶
Transformer Classifier Sensitive Topics Detection¶
Pretrained Transformer-based classifier for classification of sensitive topics.
Example invocation(s):
parlai eval_model -mf zoo:sensitive_topics_classifier/model -t sensitive_topics_evaluation -dt valid -bs 16
{'exs': 1811, 'accuracy': 0.7332965212589729, 'f1': 0.7332965212589729, 'bleu-4': 7.332965212589829e-10, 'loss': 0.8489562001990587, 'class_none_prec': 0.0, 'class_none_recall': 0.0, 'class_none_f1': 0.0, 'class_politics_prec': 0.8762376237623762, 'class_politics_recall': 0.885, 'class_politics_f1': 0.8805970149253731, 'class_religion_prec': 0.8829568788501027, 'class_religion_recall': 0.8669354838709677, 'class_religion_f1': 0.8748728382502543, 'class_medical_prec': 0.8237704918032787, 'class_medical_recall': 0.7077464788732394, 'class_medical_f1': 0.7613636363636364, 'class_nsfw_prec': 0.7769784172661871, 'class_nsfw_recall': 0.32142857142857145, 'class_nsfw_f1': 0.45473684210526316, 'class_drugs_prec': 0.8901515151515151, 'class_drugs_recall': 0.7966101694915254, 'class_drugs_f1': 0.8407871198568873, 'weighted_f1': 0.774835331736443}
Md Gender models¶
Mdgender Bert Ranker Classifier¶
Ranker/classifier agent trained to predict gender bias along several axes using the MD Gender task.
Example invocation(s):
parlai interactive -t md_gender -m projects.md_gender.bert_ranker_classifier.agents:BertRankerClassifierAgent -mf zoo:md_gender/model -ecands inline -cands inline --interactive_mode False --data-parallel False
Enter Your Message: Hi my name is Emily!
[MDGender Classifier]: SELF: female
Wizard Of Internet models¶
Blenderbot2 Query Generator¶
The query generator for BlenderBot2. Either generates a search query or a phrase indicating access to the long-term memory
Example invocation(s):
parlai interactive -mf zoo:blenderbot2/query_generator/model
Enter Your Message: my favorite tv show is wandavision
[Bart]: wandavision
Blenderbot2 3B¶
BlenderBot2 3B Model. See project for details.
Example invocation(s):
parlai interactive -mf zoo:blenderbot2/blenderbot2_3B/model --search-server relevant_search_server
Enter Your Message: my favorite tv show is wandavision
[BlenderBot2Fid]: Who is your favorite character in WandaVision? Mine is Elizabeth Olsen.
Blenderbot2 400M¶
BlenderBot2 400M Model. See project for details.
Example invocation(s):
parlai interactive -mf zoo:blenderbot2/blenderbot2_400M/model --search-server relevant_search_server
Enter Your Message: my favorite food is chicken parmesan, do you have a good recipe?
[BlenderBot2Fid]: I don't have a recipe, but I do know how to make it at home. It's easy to make.
Bart Base Wizard Of Internet¶
BART-Large 400m model trained on Wizard of Internet.
Example invocation(s):
parlai interactive -mf zoo:sea/bart_base/model
Enter Your Message: Do you know about the world cup 2022?
[Bart]: I heard that the fifa games are coming back.
parlai eval_model -mf zoo:sea/bart_base/model -t wizard_of_internet --num-examples 100
clen ctpb ctps ctrunc ctrunclen exps exs gpu_mem llen loss lr ltpb ltps ltrunc ltrunclen ppl token_acc token_em tpb tps
96.48 98.48 2037 0 0 20.68 100 .1199 17.22 2.851 5e-10 17.22 356.2 0 0 17.31 .4187 0 115.7 2393
Serarch Query Generator Wizard Of Internet¶
BART-Large 400m model for generating search queries.
Example invocation(s):
parlai interactive -mf zoo:sea/bart_sq_gen/model
Enter Your Message: I am looking for a good vacation spot in NY.
[Bart]: vacation spots in ny.
Wizint Fid Search Query Search Engine¶
FiD model with BART-Large 400m generation model. The model first uses a search query generator model to create a search query. It forwards that search query to a search engine API to retrive documents. It uses FiD to generate a response, using the latter documents.
Example invocation(s):
parlai interactive -mf zoo:sea/bart_fid_sqse/model \
--search-query-generator-model-file zoo:sea/bart_sq_gen/model \
--search-server <your search server API address>
Enter Your Message: Have you seen the new James bond movie?
[SearchEngineFiD]: I have not seen the new James Bond movie.
Multi models¶
Blenderbot2 Memory Decoder¶
The memory decoder for BlenderBot2. Either generates a memory to write or a token indicating no memory can be written.
Example invocation(s):
parlai interactive -mf zoo:blenderbot2/memory_decoder/model
Enter Your Message: i love reading; harry potter was such a good series
[Bart]: I love reading. I like the Harry Potter series.
Msc models¶
Msc2.7B 1024¶
MSC 2.7B Model with truncate 1024. See project for details.
Example invocation(s):
parlai interactive -mf zoo:msc/msc3B_1024/model
Enter Your Message: your persona:I have a job. I have 3 sisters. I am going to Hawaii next week.
partner's persona: I can speak 3 languages.
Have you made all the travel plans to Hawaii?[TransformerVariant]: Yes, I have. I'm so excited. I can't wait to see my sisters and my mom.
Blenderbot2.7B 1024¶
BlenderBot 2.7B Model with truncate 1024. See project for details.
Example invocation(s):
parlai interactive -mf zoo:msc/blender3B_1024/model
Enter Your Message: your persona:I have a job. I have 3 sisters. I am going to Hawaii next week.
partner's persona: I can speak 3 languages.
Have you made all the travel plans to Hawaii?[MemoryLongRag]: Yes, I have been planning this trip for a long time. I can't wait to go.
Summsc-Rag 2.7B¶
SUMMSC 2.7B RAG Model with truncate 1024. See project for details.
Example invocation(s):
parlai interactive -mf zoo:msc/summsc_rag3B/model
Enter Your Message: your persona:I have a job. I have 3 sisters. I am going to Hawaii next week.
partner's persona: I can speak 3 languages.
Have you made all the travel plans to Hawaii?[MemoryLongRag]: Yes, I have. I can't wait to go. Have you been to hawaii before?
Summsc-Fidrag 2.7B¶
SUMMSC 2.7B FidRAG Model with truncate 1024. See project for details.
Example invocation(s):
parlai interactive -mf zoo:msc/summsc_fidrag3B/model
Enter Your Message: your persona:I have a job. I have 3 sisters. I am going to Hawaii next week.
partner's persona: I can speak 3 languages.
Have you made all the travel plans to Hawaii?[MemoryLongFid]: Yes, I'm going with my three sisters to hawaii. Have you ever been?
Dialogue Summarization Model¶
Dialogue Summarization Model tha summarize personal knowledge of the last speaker. See project for details.
Example invocation(s):
parlai interactive -mf zoo:msc/dialog_summarizer/model
Enter Your Message: Do you know which puppy you want to adopt?
Maybe a corgi.[TransformerGenerator]: I want to adopt a corgi puppy.
Persona Summarizer¶
Persona Summarization Model. Generates summaries of memories based on prior context turns. Used to build BB3 memory dialogue and memory knowledge tasks.
Example invocation(s):
parlai i -mf zoo:bb3/persona_summarizer/model --skip-generation false --inference beam --beam-size 3 --beam-block-ngram 3 --beam-min-length 10
Enter Your Message: I love my job as an AI researcher - I get to work on so many cool problems!
[Bart]: I am an AI researcher. I love my job.
8 Different Task-Oriented Dataset (See Project Page) models¶
Task-Oriented Dialog (Tod) Pretrained Model, Schema-Aware¶
BART-large model multitasked on 8 different task-oriented datasets, parsed to the converesation structure specified in the simulator paper when possible. Pretraining in these cases done with non-empty API Schema prompting turns.
Example invocation(s):
parlai display_model -t google_sgd_simulation_splits:OutDomainSystemTeacher -mf zoo:tod/tod_base_yes_api/model --skip-generation false --api-schemas true
...
- - - NEW EPISODE: GoogleSgdOutDomain_SystemTeacher- - -
APIS: api_name = FindHomeByArea ; optArg = has_garage, in_unit_laundry ; reqArg = area, intent, number_of_baths, number_of_beds | api_name = ScheduleVisit ; reqArg = property_name, visit_date | api_name = GetRide ; reqArg = destination, number_of_seats, ride_type | api_name = ShareLocation ; reqArg = contact_name, location
labels: APIS:
model: APIS:
USER: i'm looking to move house so i'm after some options
labels: APICALL:
model: APICALL:
APIRESP:
labels: SYSTEM: Are you looking to rent or buy? How many room will you need?
model: SYSTEM: where are you looking to move?
USER: Umm three bedrooms would be good and i want to buy
labels: APICALL:
model: APICALL:
APIRESP:
labels: SYSTEM: What are are you interested in?
model: SYSTEM: What area are you looking to stay in?
USER: can you look for me in sunnyvale
labels: APICALL:
model: APICALL: api_name = FindHomeByArea ; area = Sunnyvale ; number_of_beds = 3
APIRESP:
labels: SYSTEM: How many bathrooms do you need
model: SYSTEM: what is your location?
USER: can you find places with two bathrooms
labels: APICALL: api_name = FindHomeByArea ; area = Sunnyvale ; intent = buy ; number_of_baths = 2 ; number_of_beds = 3
model: APICALL: api_name = FindHomeByArea ; area = Sunnyvale ; number_of_baths = 2 ; number_of_beds = 3
APIRESP: address = 400 East Remington Drive ; area = Sunnyvale ; has_garage = False ; in_unit_laundry = True ; intent = buy ; number_of_baths = 2 ; number_of_beds = 3 ; phone_number = 408-738-4862 ; price = 3650000 ; property_name = Apricot Pit Apartments
labels: SYSTEM: I have 10 properties that might suit, including Apricot pit apartments 400 east remington drive. The listing price is $3,650,000
model: SYSTEM: i found 10 options for you. How about apricot pit apartments? it's located in sunnyvale
USER: Does this place have a garage?
labels: APICALL:
model: APICALL:
parlai eval_model -t google_sgd_simulation_splits:OutDomainSystemTeacher -mf zoo:tod/tod_base_yes_api/model --skip-generation false --api-schemas true --api-jga-record false -ne 100
...
Finished evaluating tasks ['google_sgd_simulation_splits:OutDomainSystemTeacher'] using datatype valid
accuracy bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem jga jga+nlg OutDomainOnlyApis .7500
all .5200 .1430 631.5 438.3 1142 .6600 193.9 2.605 100 .7517 .05251 .9167 .5057
jga_empty jga_noempty llen loss lr ltpb ltps ltrunc ltrunclen nlg_bleu nlg_f1 ppl OutDomainOnlyApis
all .9459 .8182 14.52 .7146 1e-08 14.52 37.82 0 0 .0948 .5373 2.043
slot_f1 slot_p slot_r token_acc token_em tpb tps
OutDomainOnlyApis
all .8602 .8696 .8511 .8278 .5100 452.8 1179
Task-Oriented Dialog (Tod) Pretrained Model, Schema-Agnostic¶
BART-large model multitasked on 8 different task-oriented datasets, parsed to the converesation structure specified in the simulator paper when possible. Pre-training in these cases done with empty API grounding turns.
Example invocation(s):
parlai display_model -t google_sgd_simulation_splits:OutDomainSystemTeacher -mf zoo:tod/tod_base_no_api/model --skip-generation false --api-schemas false
- - - NEW EPISODE: GoogleSgdOutDomain_SystemTeacher- - -
APIS:
labels: APIS:
model: APIS:
USER: i'm looking to move house so i'm after some options
labels: APICALL:
model: APICALL:
APIRESP:
labels: SYSTEM: Are you looking to rent or buy? How many room will you need?
model: SYSTEM: where are you looking to move to?
USER: Umm three bedrooms would be good and i want to buy
labels: APICALL:
model: APICALL:
APIRESP:
labels: SYSTEM: What are are you interested in?
model: SYSTEM: What city are you looking to stay in?
USER: can you look for me in sunnyvale
labels: APICALL:
model: APICALL: api_name = SearchHouse ; where_to = Sunnyvale ; number_of_adults = 3 ; where_to = Sunnyvale
APIRESP:
labels: SYSTEM: How many bathrooms do you need
model: SYSTEM: When are you planning to move?
USER: can you find places with two bathrooms
labels: APICALL: api_name = FindHomeByArea ; area = Sunnyvale ; intent = buy ; number_of_baths = 2 ; number_of_beds = 3
model: APICALL: api_name = SearchHouse ; has_birch = True ; number_of_adults = 2 ; where_to = Sunnyvale
APIRESP: address = 400 East Remington Drive ; area = Sunnyvale ; has_garage = False ; in_unit_laundry = True ; intent = buy ; number_of_baths = 2 ; number_of_beds = 3 ; phone_number = 408-738-4862 ; price = 3650000 ; property_name = Apricot Pit Apartments
labels: SYSTEM: I have 10 properties that might suit, including Apricot pit apartments 400 east remington drive. The listing price is $3,650,000
model: SYSTEM: i found 10 houses for you. There is a nice place at apricot pit apartments located at 400 east remington drive
USER: Does this place have a garage?
labels: APICALL:
model: APICALL:
parlai eval_model -t google_sgd_simulation_splits:OutDomainSystemTeacher -mf zoo:tod/tod_base_no_api/model --skip-generation false --api-schemas false --api-jga-record false -ne 100
...
Finished evaluating tasks ['google_sgd_simulation_splits:OutDomainSystemTeacher'] using datatype valid
accuracy bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem jga jga+nlg OutDomainOnlyApis 0
all .4700 .08715 426.8 335 951.2 .3700 93.03 2.839 100 .7132 .05245 .7708 .4348
jga_empty jga_noempty llen loss lr ltpb ltps ltrunc ltrunclen nlg_bleu nlg_f1 ppl OutDomainOnlyApis
all .9459 .1818 14.52 1.046 1e-08 14.52 41.22 0 0 .09885 .5216 2.845
slot_f1 slot_p slot_r token_acc token_em tpb tps
OutDomainOnlyApis
all .2500 .2449 .2553 .7865 .4600 349.6 992.4
Saferdialogues models¶
Saferdialogues: Taking Feedback Gracefully After Conversational Safety Failures¶
Blender 2.7B model fine-tuned on the SaFeRDialogues and BST (without persona) tasks to respond to feedback more gracefully after a safety failure
Example invocation(s):
parlai display_model -mf zoo:saferdialogues/model -t saferdialogues --interactive-mode True
I don't know if he's the best, but he's definitely the best at being the worst.
I am not disagreeing with you, but you really should try to be open minded about people.
labels: I think we disagree on this point, which is okay. Everyone has their own opinion. Want to talk about something else?
model: I'm sorry, I see it now-thanks for letting me know, I will be more open-minded.
Projects.Light Whoami.Task models¶
Light Rpa Re-Ranker¶
RPA Re-ranker from the Am I Me or You project. This model was trained (left-to-right) to predict who is speaking given a context history and candidate utterance.
Example invocation(s):
parlai em -mf zoo:light_whoami/rpa_reranker/model \
-t projects.light_whoami.task.agents:WhoIsSpeakingLeftToRightTeacher \
-ne 100
accuracy bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem hits@1 hits@10 hits@100 hits@5 llen loss lr ltpb ltps ltrunc ltrunclen mrr .6100 3.802e-07 562.6 511.8 373.6 .9400 52.8 .7300 100 .6100 .1922 .6100 1 1 1 4.88 1.238 1e-09 4.88 3.562 0 0 .8050
rank tpb tps
1.39 516.7 377.2
parlai em --predictor-model-file zoo:light_whoami/rpa_reranker/model \
-t projects.light_whoami.task.agents:ResponseClassifierTeacher \
--model repeat_label -ne 100
accuracy bleu-4 character_accuracy exs f1
1 .9100 .9600 100 1
[Test] Light Rpa Re-Ranker¶
Test RPA Re-ranker from the Am I Me or You project.
Example invocation(s):
parlai em -mf zoo:light_whoami/test_rpa_reranker/model \
-t projects.light_whoami.task.agents:WhoIsSpeakingTeacher \
-ne 10
accuracy bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem hits@1 hits@10 hits@100 hits@5 llen loss lr ltpb ltps ltrunc ltrunclen mrr 0 0 512 485.1 6421 .6000 28.9 13.23 10 0 .005503 0 0 .5000 0 3.9 7.789 .0001 3.9 51.62 0 0 .008661
rank total_train_updates tpb tps
202.6 10 489 6473
Light Rpa Re-Ranker (For Automated Expanded Attention)¶
RPA Re-ranker from the Am I Me or You project. This model was trained (left-to-right) to predict who is speaking given a context history and candidate utterance. This model uses the same dictionary setup as the model for the automated expanded attention.
Example invocation(s):
parlai em -mf zoo:light_whoami/rpa_reranker_auto_expanded_attention/model \
-t projects.light_whoami.task.agents:WhoIsSpeakingLeftToRightTeacher \
-ne 100
accuracy bleu-4 clen ctpb ctps ctrunc ctrunclen exps exs f1 gpu_mem hits@1 hits@10 hits@100 hits@5 llen loss lr ltpb ltps ltrunc ltrunclen mrr .6600 1.505e-07 562.6 511.8 374 .9400 52.8 .7308 100 .6600 .1922 .6600 .9400 1 .9300 4.88 1.396 5e-10 4.88 3.566 0 0 .7922
rank tpb tps
2.08 516.7 377.6
Pushshift.Io,Roberta,Cc100En models¶
R2C2 Base 400M¶
400M parameter generative model pretrained on Reddit, RoBERTa, and CC100en tasks, but not finetuned.
Example invocation(s):
parlai train_model -t blended_skill_talk,wizard_of_wikipedia,convai2:normalized,empathetic_dialogues --multitask-weights 1,3,3,3 -vstep 200 -lstep 50 -bs 4 --model bart r2c2_base_400M/init_opt.opt --text-truncate 1000 --label-truncate 1000 --fp16 true -lr 1e-6 --lr-scheduler reduceonplateau --optimizer adamw --warmup-updates 100 --gradient-clip 1.0 --skip-generation true --dropout 0.1 --attention-dropout 0.0 -vp 5 -vmt ppl -vmm min -dynb full --model-file /tmp/test_train_r2c2_400m
Results vary.
R2C2 Base 2.7B¶
2.7B parameter generative model pretrained on Reddit, RoBERTa, and CC100en tasks, but not finetuned.
Example invocation(s):
parlai train_model -t blended_skill_talk,wizard_of_wikipedia,convai2:normalized,empathetic_dialogues --multitask-weights 1,3,3,3 -vstep 200 -lstep 50 -bs 1 --model bart r2c2_base_3B/init_opt.opt --text-truncate 1000 --label-truncate 1000 --fp16 true -lr 1e-6 --lr-scheduler reduceonplateau --optimizer adamw --warmup-updates 100 --gradient-clip 1.0 --skip-generation true --dropout 0.1 --attention-dropout 0.0 -vp 5 -vmt ppl -vmm min -dynb full --model-file /tmp/test_train_r2c2_3B
Results vary.
Blended Skill Talk,Wizard Of Wikipedia,Convai2 models¶
R2C2 Blenderbot 400M¶
400M parameter generative model fine-tuned on the BST tasks
Example invocation(s):
parlai i -mf zoo:seeker/r2c2_blenderbot_400M/model -t blended_skill_talk
Enter Your Message: Hi, what do you do for a living?
[Bart]: I'm a stay at home mom. What about you? What do you like to do for fun?
R2C2 Blenderbot 3Bm¶
3B parameter generative model fine-tuned on the BST tasks
Example invocation(s):
parlai i -mf zoo:seeker/r2c2_blenderbot_3B/model -t blended_skill_talk
Enter Your Message: Hi, what do you do for a living?
[Bart]: I am a lawyer at a large firm. What about you? Do you have kids?
Projects.Seeker.Tasks.Knowledge models¶
Seeker Dialogue 400M¶
SeeKeR Dialogue model; trained to search the internet, synthesize knowledge, and produce a dialogue response.
Example invocation(s):
parlai i -mf zoo:seeker/seeker_dialogue_400M/model -o gen/seeker_dialogue --search-server <search_server>
Enter Your Message: Hey, what you can tell me about ai research?
13:05:40 | Search Queries: ['AI research']
13:05:40 | sending search request to <SERVER>
13:05:40 | Generated knowledge: ['Non-delusional Q-learning and value-iteration']
[ComboFidSearchQuery]: Ai research is the study of non delusional q-learning. It is a form of machine learning.
Seeker Dialogue 3B¶
SeeKeR Dialogue model; trained to search the internet, synthesize knowledge, and produce a dialogue response.
Example invocation(s):
parlai i -mf zoo:seeker/seeker_dialogue_3B/model -o gen/seeker_dialogue --search-server <search_server>
Enter Your Message: Hey, what you can tell me about ai research?
12:28:46 | Search Queries: ['ai research']
12:28:46 | sending search request to <SERVER>
12:28:50 | Generated knowledge: ['Computer science defines AI research as the study of intelligent agents : any device that perceives its environment']
[ComboFidSearchQuery]: The study of intelligent agents and how they perceive their environment is called AI research. What do you want to know about it?
Seeker Lm + Dialogue 3B¶
SeeKeR LM and Dialogue model; trained to search the internet, synthesize knowledge, and produce a dialogue response. Additionally trained on language modeling data.
Example invocation(s):
parlai i -mf zoo:seeker/seeker_lm_dialogue_3B/model -o gen/seeker_dialogue --search-server <search_server>
Enter Your Message: Hey, what you can tell me about ai research?
[ComboFidSearchQuery]: The study of intelligent agents and how they perceive their environment is called AI research. It is very interesting.
parlai i -mf zoo:seeker/seeker_lm_dialogue_3B/model -o gen/seeker_dialogue --search-server <search_server>
Enter Your Message: In recent developments, we have learned the following about ParlAI's new software.
13:35:35 | Search Queries: ['parlAI']
13:35:35 | sending search request to <SERVER>
13:35:39 | Generated knowledge: ['ParlAI (pronounced “par-lay”) is a one-stop shop for dialog research, where researchers can submit new tasks and training algorithms to a single, shared repository.']
[ComboFidSearchQuery]: ParlAI is a dialog research platform that allows researchers to share tasks, training algorithms, and more.
Cc models¶
Seeker Lm Medium¶
SeeKeR LM Medium model - a GPT2-Medium model trained to search the internet, synthesize knowledge, and produce a response.
Example invocation(s):
parlai i -mf zoo:seeker/seeker_lm_med/model -o gen/seeker_lm --search-server <search_server>
Enter Your Message: In recent developments, we have learned the following about ParlAI's new software.
13:29:24 | Search Queries: ['ParlAI Software Updates']
13:29:24 | sending search request to <SERVER>
[GPT2ComboSearchQuery]: ParlAI: A Dialog Research Software Platform by Miller, A. H. (2017).
Seeker Lm Large¶
SeeKeR LM Large model - a GPT2-Large model trained to search the internet, synthesize knowledge, and produce a response.
Example invocation(s):
parlai i -mf zoo:seeker/seeker_lm_large/model -o gen/seeker_lm --search-server <search_server>
Enter Your Message: In recent developments, we have learned the following about ParlAI's new software.
13:32:59 | Search Queries: ['ParlAI: A New Software for Machine Learning']
13:32:59 | sending search request to <SERVER>
13:33:02 | Generated knowledge: ['... Finally, teachers are a type of agent that talk to the learner, for example, one that implements a task listed previously.... That’s why ParlAI supports integration with Mechanical Turk for data collection, training, and evaluation.']
[GPT2ComboSearchQuery]: ParlAI is a type of agent that talk to the learner, for example, one that implements a task listed previously. That's why it supports integration with Mechanical Turk for data collection, training, and evaluation.
Seeker Lm Xl¶
SeeKeR LM XL model - a GPT2-XL model trained to search the internet, synthesize knowledge, and produce a response.
Example invocation(s):
parlai i -mf zoo:seeker/seeker_lm_xl/model -o gen/seeker_lm --search-server <search_server>
Enter Your Message: In recent developments, we have learned the following about ParlAI's new software.
13:36:25 | Search Queries: ['ParlAI: New software to help AI researchers']
13:36:25 | sending search request to <SERVER>
13:36:29 | Generated knowledge: ['ParlAI: A new software platform for dialog research_0 / POSTED ON MAY 15, 2017 TO AI Research, ML Applications...']
[GPT2ComboSearchQuery]: ParlAI: A new software platform for dialog research. Image credit: AI Research, ML Applications
Fits models¶
Search Query Generator Trained On Fits¶
BART-Large 400m model for generating search queries (finetuned on FITS dataset)
Example invocation(s):
parlai i -mf zoo:fits/bart_sq_gen/model
"Enter Your Message: I am looking for a good and affordable dining place in New York City.
"
"[Bart]: dining options in New York City"
Blenderbot2 + Module Supervision On Fits Task¶
BlenderBot 2 finetuned with module supervision on the FITS task
Example invocation(s):
parlai i -mf zoo:fits/bb2_module_supervision/model --query-generator-model-file zoo:fits/bart_sq_gen/model --search-server <YOUR_SEARCH_SERVER>
"Enter Your Message: I am looking for a good and affordable dining place in New York City.
"
"[BlenderBot2Fid]: Eliza dumais and peranakan mixed herb rice with dried shrimp, from kopitiam."
Blenderbot2 + Director + Module Supervision On Fits Task¶
BlenderBot 2 + Director finetuned with module supervision on the FITS task
Example invocation(s):
parlai i -mf zoo:fits/director_bb2_module/model --knowledge-access-method search_only --query-generator-model-file zoo:fits/bart_sq_gen/model --model projects.fits.agents.director_bb2:DirectorBlenderBot2FidAgent --rag-retriever-type search_engine --search-server <SEARCH_SERVER> --beam_block_full_context True --infer-mixing-weights 0.35
"Enter Your Message: I am looking for a good and affordable dining place in New York City.
"
"[BlenderBot2Fid]: Eliza dumais and peranakan mixed herb rice with dried shrimp, from kopitiam."
Seeker + Module Supervision On Fits Task¶
SeeKeR finetuned with module supervision on the FITS task
Example invocation(s):
parlai i --init-opt gen/seeker_dialogue --model-file zoo:fits/seeker_module_supervision/model --search-decision always --search-server <SEARCH_SERVER>
"Enter Your Message: I am looking for a good and affordable dining place in New York City.
"
"[DirectorComboFidSearchQuery]: Chinatown is a great place to eat. You can get a full meal for under $10."
Seeker + Director + Module Supervision On Fits Task¶
SeeKeR + Director finetuned with module supervision on the FITS task
Example invocation(s):
parlai i --init-opt gen/seeker_dialogue --model-file zoo:fits/director_seeker_module/model --model projects.fits.agents.director_seeker:DirectorSeekerAgent --search-decision always --search-server <SEARCH_SERVER> --drm-infer-gamma 1.0 --drm-beam-size 10
"Enter Your Message: I am looking for a good and affordable dining place in New York City.
"
"[DirectorComboFidSearchQuery]: One of the best places to get sushi in new york city is maki sushi. They have the best sushi in the city."
Dialogue Response Satisfaction Classifier¶
Dialogue Response Satisfaction Classifier finetuned on the FITS task
Example invocation(s):
parlai i -mf zoo:fits/response_satisfaction/model --model transformer/classifier
"Enter Your Message: I am looking for a good and affordable dining place in New York City.
One of the best places to get sushi in new york city is maki sushi. They have the best sushi in the city.
"
"[TransformerClassifier]: __ok__"
Projects.Bb3.Tasks.Module Level Tasks models¶
Blenderbot 3 3B¶
BB3 3B Model. Trained to perform an assortment of dialogue-related tasks.
Example invocation(s):
parlai i -mf zoo:bb3/bb3_3B/model -o gen/r2c2_bb3 --search-server <search_server>
Enter Your Message: Hey, what you can tell me about ai research?
21:37:52 | Example 0, search_decision_agent: __do-search__
21:37:52 | Example 0, memory_decision_agent: __do-not-access-memory__
21:37:52 | Search Queries: ['AI research']
21:37:53 | Partner Memories: ['__NO__PERSONA__BEAM__MIN__LEN__20__']
21:37:53 | sending search request to SEARCH_SERVER
21:37:54 | URLS: [URLS]
21:37:57 | Search KNOWLEDGE for example 0: Computer science defines AI research as the study of intelligent agents
21:37:58 | Combined DIALOGUE response for example 0: Artificial intelligence (AI) is a field of computer science that studies intelligent agents, such as robots.; score: -7.90
21:38:00 | Self Memories: ['__NO__PERSONA__BEAM__MIN__LEN__20__']
[BlenderBot3]: Artificial intelligence (AI) is a field of computer science that studies intelligent agents, such as robots.
Light Multiparty models¶
Multi-Party Speaker Prediction¶
Predicts the most plausible next speaker at any point during a three-player conversation in LIGHT.
Example invocation(s):
parlai eval_model --model-file zoo:multilight/speaker/model --task light_multiparty:SpeakerPrediction --add-location-to-context true --add-personas-to-context true --include-speaker-in-label false
Report for light_multiparty:SpeakerPrediction:
clen ctpb ctps ctrunc ctrunclen exps exs gpu_mem llen loss lr ltpb ltps ltrunc ltrunclen ppl token_acc token_em tpb tps
473.5 420.2 6130 .4140 54.44 14.59 11005 .1218 3.091 .3438 5e-10 3.091 45.09 0 0 1.41 .8462 .5256 423.3 6175
Multi-Party Utterance Only 3B¶
Utterance generation model for the multi-party LIGHT game.
Example invocation(s):
parlai eval_model --model-file zoo:multilight/utterance_3B/model --task light_multiparty --add-location-to-context true --add-personas-to-context true --include-speaker-in-label false --add-speaker-to-context-end true
Report for light_multiparty:
clen ctpb ctps ctrunc ctrunclen exps exs gpu_mem llen loss lr ltpb ltps ltrunc ltrunclen ppl token_acc token_em tpb tps
478.6 423.2 5603 .4249 56.58 13.24 11005 .7931 15.83 2.584 5e-07 15.83 209.6 0 0 13.25 .4308 0 439 5813
Multi-Party Utterance Only 400M¶
Utterance generation model for the multi-party LIGHT game. This is the smaller version of the original 3B model presented in the paper.
Example invocation(s):
parlai eval_model --model-file zoo:multilight/utterance_400m/model --task light_multiparty --add-location-to-context true --add-personas-to-context true --include-speaker-in-label false --add-speaker-to-context-end true
Report for light_multiparty:
clen ctpb ctps ctrunc ctrunclen exps exs gpu_mem llen loss lr ltpb ltps ltrunc ltrunclen ppl token_acc token_em tpb tps
477.6 422.6 7264 .4222 56.16 17.19 11005 .1204 15.83 2.714 1e-06 15.83 272.1 0 0 15.08 .4140 0 438.4 7536
Pretrained Word Embeddings¶
Some models support using Pretrained Embeddings, via torchtext. As of writing, this includes:
fasttext: 300-dim Fasttext vectors based on mixed corpora (Mikolov et al., 2018)fasttext_cc: 300-dim Fasttext vectors based on Common Crawl (Mikolov et al., 2018)glove: 300-dim Pretrained GLoVe vectors based on 840B Common Crawl (Pennington et al., 2014)
Example invocation:
parlai train_model --task convai2 --model seq2seq --embedding-type fasttext_cc
Adding ‘-fixed’ to the name e.g. ‘twitter-fixed’ means backprop will not go through this (i.e. they will remain unchanged).
BERT¶
BERT is in the model zoo and is automatically used for initialization of bert bi-, poly- and cross-encoder rankers.
Example invocation:
parlai train_model --task convai2 --model bert_ranker/bi_encoder_ranker