issues
search
opentensor
/
validators
Repository for bittensor validators
https://www.bittensor.com/
MIT License
14
stars
9
forks
source link
Dpo penalty update
#138
Closed
Eugene-hu
closed
1 year ago
Eugene-hu
commented
1 year ago
Adds additional checks for empty strings and smaller completions
Adds a penalty for repeated tokens akin to huggingface's repeat penalty during generation (
https://github.com/huggingface/transformers/blob/v4.32.0/src/transformers/generation/logits_process.py#L279
)
The penalty will deter repeated tokens with subsequently heavier penalties each time it occurs.