taprosoft / llm_finetuning Goto Github PK

Is it possible to do raw-text training (not instruction style - just feeding full text of literature)?

If so could you suggest a template to do this?

requirements.txt has git+https://github.com/huggingface/peft.git

cuda_quant_requirements.txt has peft==0.4.0

when I try and install both of these, I see the following

Collecting git+https://github.com/huggingface/peft.git (from -r requirements.txt (line 1))
  Cloning https://github.com/huggingface/peft.git to /tmp/pip-req-build-n5akfjww
  Running command git clone --filter=blob:none --quiet https://github.com/huggingface/peft.git /tmp/pip-req-build-n5akfjww
  Resolved https://github.com/huggingface/peft.git to commit ad8f7cb59ee7ca4b9ca1c9048711038ac36b31b8
  Installing build dependencies ... done
  Getting requirements to build wheel ... done
  Preparing metadata (pyproject.toml) ... done
ERROR: Cannot install peft 0.11.2.dev0 (from git+https://github.com/huggingface/peft.git) and peft==0.4.0 because these package versions have conflicting dependencies.

The conflict is caused by:
    The user requested peft 0.11.2.dev0 (from git+https://github.com/huggingface/peft.git)
    The user requested peft==0.4.0

To fix this you could try to:
1. loosen the range of package versions you've specified
2. remove package versions to allow pip attempt to solve the dependency conflict

ERROR: ResolutionImpossible: for help visit https://pip.pypa.io/en/latest/topics/dependency-resolution/#dealing-with-dependency-conflicts

I’m trying to finetune my gptq pretrained model via your framework, but I found that gpt+lora seems to make the model worse.

During the initialization phase of training, I met the warning:
The safetensors archive passed at gptq_model-4bit-128g.safetensors does not contain metadata. Make sure to save your model with the save_pretrained method. Defaulting to 'pt' metadata.

hope you can give me some advice, thanks!

Could you provide a sample template for training a character chat model (like "USER: Hello\nASSISTANT: Hi") except back and forth?

It'd be easy to modify the existing template samples for a single USER input followed by one ASSISTANT output but how can the sharegpt:chat conversation format be implemented (like axolotl supports) where there is an array of chat?

{"conversations": [
  {"from": "human", "value": "hi"},
  {"from": "gpt", "value": "hi"},
  {"from": "human", "value": "how are you"},
  {"from": "gpt", "value": "good"}
]}

I formatted my data as shareGPT but it keeps throwing up this error.

Failed to read file '/workspace/llm_finetuning/datasets/dataset.json' with error <class 'pyarrow.lib.ArrowInvalid'>: Could not convert 'USER' with type str: tried to convert to int64
Traceback (most recent call last):
File "/usr/local/lib/python3.10/dist-packages/datasets/builder.py", line 1879, in _prepare_split_single
for _, table in generator:
File "/usr/local/lib/python3.10/dist-packages/datasets/packaged_modules/json/json.py", line 156, in _generate_tables
raise ValueError(f"Not able to read records in the JSON file at {file}.") from None
ValueError: Not able to read records in the JSON file at /workspace/llm_finetuning/datasets/dataset.json.

petf@git+https://github.com/huggingface/peft.git
Should be
peft@git+https://github.com/huggingface/peft.git