togethercomputer / OpenChatKit

Apache License 2.0
9k stars 1.01k forks source link

TODO: Better Documentation for hyperparameters and fine tuning #10

Open zhangce opened 1 year ago

zhangce commented 1 year ago
  1. How to reproduce HF App's behavior (hyper-parameters, exact prompt)
  2. How to fine-tune models given a new corpus + things to be careful about and best practice
pdtgct commented 1 year ago

Here is an example generation_config.json:

{
  "_from_model_config": true,
  "bos_token_id": 0,
  "eos_token_id": 0,
  "pad_token_id": 1,
  "max_new_tokens": 128,
  "min_new_tokens": 1,
  "penalty_alpha": null,
  "repetition_penalty": 1.0,
  "do_sample": true,
  "temperature": 0.6,
  "top_k": 50,
  "top_p": 0.25,
  "early_stopping": true,
  "no_repeat_ngram_size": 2,
  "length_penalty": 1.4,
  "num_beams": 1,
  "use_cache": true,
  "transformers_version": "4.26.0"
}
samarthsarin commented 1 year ago

@csris Any update on the ipynb fine tuning documentation on the custom data article?

csris commented 1 year ago

@zhangce, any updates on this?