What's the difference between llm-blender/PairRM and llm-blender/pair-ranker?
#3
by
nefelibata-mu
- opened
I'm wondering the difference between llm-blender/PairRM and llm-blender/pair-ranker. Hope to get an answer.
Their model architecture is the same. The main difference lies in the training data and context length.
- For the training data:
pair-rankeris the ranker trained on llm-blender/mix-instruct dataset, producing the results reported in the llm-blender paper.PairRMis trained onopenai/summarize_from_feedback,lmsys/chatbot_arena_conversations, etc. which does not containmix-instructdata. (see in PairRM README)
- For the context length
There is a simple table comparing the this. The main difference is thatpair-rankerconstrains the source and candidate lengths to be shorter than 128, whilePairRMcan extend that constraint to 1224 and 412 respectively.
Overall, you can consider PairRM a more powerful version of llm-blender/pair-ranker.
Thank you very much for your patient answer.