lm-sys / lm-sys/FastChat

Confusions about prompt in training stage

Open
#2,574 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
39.5k
Forks
4.8k
PR merge metrics
No merged PRs in 30d

Description

Hi, teams. Congratulations on your great work!
In finetuning code, I notice the program gets "vicuna" conversation template. It will turn samples in datasets to "It's a chat between user and assistant.... USER:xxx. ASSISTANT:xxx" and mark the response part to let model update its params.
I have two confusions about this training stage prompt.

  1. Does this prompt make a difference? Can we change prompt as we want?
  2. I notice this finetuning code is for llama model. If I change to another model, can I still use this finetuning code? I mean can I finetune other models by running train.py instead of train_mem.py which is specific to llama? I have this confusion because I think train.py is not specific to llama model except it uses "vicuna" conversation. If this prompt is a general one, we can use this finetuning code to any model in my thought.

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Read train.py and train_mem.py alongside the finetuning code that selects the vicuna conversation template. Determine which parts depend on the LLaMA model and whether prompt and model compatibility are documented; done means recording clear answers for changing the prompt and using other model architectures.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
documentation, machine-learning
Issue type
Documentation
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.