Skip to content

Latest commit

 

History

History
121 lines (77 loc) · 4.38 KB

README.md

File metadata and controls

121 lines (77 loc) · 4.38 KB
Raw

RLHF (Reinforcement learning from human feedback) training and usage

This repo comprises a library and modules to allow easy trainig, deploying, and usage of fine-tuned models using Reinforcement Learning from Human Feedback (RLHF), incorporating a reward model alongside Proximal Policy Optimization (PPO) techniques.

By default models are loaded to be quantized with PEFT / LoRA (along with other optimization techniques) to be able to run this more easily on consumer hardware, though this can be turned off.

RLHF leverages human feedback to guide the learning process, enabling models to improve their predictions over time based on evaluative input. This approach integrates a rewards model to quantitatively assess the quality of model outputs, using these assessments to inform the PPO algorithm's optimization process.

There is a CLI to train and use the models.

Note: This library is not going to be made available via pip, as this is just a starting point made open-source, and not meant for production use.

Pre-requisites

Make and activate virtual env

python -m venv venv

source venv/bin/activate

Get CUDA drivers

Install CUDA drivers.

Install the appropriate version of PyTorch from the website: https://pytorch.org/get-started/locally/

For example, for Cuda 12.1:

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

Install the other packages

Install requirements.

pip install -r requirements.txt

These libraries were installed on a Windows machine, so they may require different sources / versions. If you run into errors, you'll have to install the right ones manually.

Other packages used: bitsandbytes, TRL, accelerate, transformers

Make your env file

cp .env.sample .env

Get your Hugging Face user access token to download models from their hub: https://huggingface.co/docs/hub/en/security-tokens, and add that to HUGGING_FACE_HUB_TOKEN env var.

RLHF Library usage

Train a fine-tuned PPO model:

from .nlp.main import NLP
nlp = NLP(
    default_model_name="mistralai/Mistral-7B-v0.1",
    rewards_model_name="facebook/opt-350m",
    models_dir="./nlp/models",
    use_lora=True,
)

ppo_name = nlp.ppo.train_ppo(
    model_name=nlp.model_name,
    ppo_model_name="my_ppo_model",
    dataset_name="ppo_finetune.csv",
    local_only=False,
)

print("Trained PPO model at: " + ppo_name + ".")

Evaluate / predict from a PPO model

from .nlp.main import NLP
ppo_model_name = "my_ppo_model_name"
query = "The movie was really"
nlp = NLP(
    models_dir="./nlp/models",
    load_model=False,
    use_lora=True,
)
nlp.ppo.load_ppo_model(ppo_model_name, local_only=True,     
use_lora=True)
response = nlp.generate_response_from_ppo(query)
print(response)

RLHF CLI usage

Train a fine-tuned PPO model:

Example:

python cli.py train ppo "ppo_finetune.csv" "meta-llama/Llama-2-7b-chat-hf" "facebook/opt-350m" "Anthropic/hh-rlhf" "ppo-model-finetuned"

  • First arg (after specifying type of CLI, train and ppo): name of dataset to fine-tune PPO model from (should be a local CSV file inside /data folder or the directory passed as data_dir in PPO class, with a column called query containing texts / prompts)
  • Second arg: name of base model to fine-tune (can be local or from HF hub)
  • Third arg: name of the rewards model to use for PPO training (can be local or from HF hub).
  • Fourth arg: name of dataset to use for rewards model training (can be local or from HB hub)
  • Fith arg: (optional): name of fine-tuned model to save. Defaults to name of base model inside the folder /ppo in /models.

Evaluate / make prediction from a HF or local model:

Example:

python cli.py eval model "meta-llama/Llama-2-7b-chat-hf" "The movie was really"

  • First arg (after specifying type of CLI, eval and model): Name of model (can be local or from HB hub) to make prediction from
  • Second arg: The query to make a prediction from

Evaluate / make prediction from a local fine-tuned PPO model (saved inside /models/ppo)

Example:

python cli.py eval ppo "my-fine-tuned-model-ppo" "The move was really"

  • First arg (after specifying type of CLI, eval and ppo): Name of model (can be local or from HB hub) to make prediction from
  • Second arg: The query to make a prediction from

Linting

black .