autolabel
autolabel copied to clipboard
[Feature Request]: Support logit_bias to constrain LLM outputs
Is your feature request related to a problem? Please describe. LLM might sometimes have trouble following output format instructions
Describe the solution you'd like logit_bias: https://twitter.com/AAAzzam/status/1669753725093654565
We can start with support for classification and entity matching tasks
thanks @neildhruva for the suggestion!
logit_bias is an OpenAI only feature at the moment. For huggingface equivalent, we can use this: https://huggingface.co/docs/transformers/main/en/internal/generation_utils#transformers.SequenceBiasLogitsProcessor