How mchatzakis/chatmgl works
Project type and purpose
• This is an ML research/engineering pipeline: a large-language model (ChatMGL) fine-tuned on data-science Q&A.
• It includes data preparation, supervised & RLHF training, multiple reward models, evaluation, and a simple serving interface.
File structure
Top-level directories/files:
• data/ – raw and split JSON datasets (MATH, MathQA, Reward, StackExchange, interactions)
• dataset/ – combined supervised and RLHF dataset JSONs used during training
• models/ – pre-trained generative base (BART) and reward model checkpoints (custom BERT, DistilBERT, GPT-2)
• src/
– src/dataset/ (dataset classes: RLHF_dataset.py, generative_dataset.py)
– src/utils/ (utility functions)
– src/generative_model.py (GenerativeModel class: load and generate)
– src/reward_model.py (wrapper for reward models)
– src/gen_script_chatMGL.py (CLI for batch generation)
– src/RLHF_pipeline.ipynb & notebooks/ (pipeline and eval experiments)
• scripts: get_chatMGL.sh (download model via gdown), run.sh (invoke generation)
• config/deps: requirements.txt, python.txt
README
• Describes installation, model download via Google Drive (gdown), generation workflow (run.sh → gen_script_chatMGL.py), usage of GenerativeModel API, seed control for reproducibility.
• Mentions datasets in repo and links to full model.