Opt Presets

Opt presets are a way to provide multiple options on the command line as shorthand. Opt presets are bundled with ParlAI and may be used by simply invoking the -o preset_name option within any ParlAI command.

You may also define your own options by placing them in ~/.parlai/opt_presets/. For example, creating ~/.parlai/opt_presets/myfolder/mypreset.opt allows you to invoke it via -o myfolder/mypreset. These preset files are simple json files containing a dictionary of files. For example:

    "inference": "beam",
    "beam_size": 10,

List of presets

The following is a list of all options presets bundled with the latest version of ParlAI.

Preset name




Architecture parameters (number layers, etc.) for BART-Large. See Lewis et. al. (2019

--activation gelu --attention-dropout 0.0 --dict-file zoo:bart/bart_large/model.dict --dict-tokenizer gpt2 --dropout 0.1 --embedding-size 1024 --embeddings-scale False --ffn-size 4096 --force-fp16-tokens True --fp16 True --init-model zoo:bart/bart_large/model --learn-positional-embeddings True --model bart --n-decoder-layers 12 --n-encoder-layers 12 --n-heads 16 --n-positions 1024 --variant bart


Architecture parameters (number layers, etc) for BlenderBot 3B. See Roller et al. (2020)

--activation gelu --attention-dropout 0 --embedding-size 2560 --ffn-size 10240 --label-truncate 128 --model transformer/generator --n-decoder-layers 24 --n-encoder-layers 2 --n-heads 32 --n-positions 128 --relu-dropout 0 --text-truncate 128 --truncate 128 --variant prelayernorm


Beam search parameters for BlenderBot. SeeRoller et al. (2020)

--beam-context-block-ngram 3 --beam-block-ngram 3 --beam-size 10 --inference beam --beam-min-length 20 --beam-block-full-context False


Inference parameters for the Sample & Rank procedure of Meena. See Adiwardana et al. (2020).

--beam-size 20 --inference topk --topk 40