FineTuning Mistral 7B Repository

Overview

This repository is for fine-tuning the Mistral 7B model using qLoRa. It includes scripts for preprocessing raw data and generating specific prompts from llama-13b from given texts

Files

finetune.py

• Fine-tuning the Mistral7B model
• Features:
  • Training on 4x NVIDIA 4090 GPUs
  • Utilises BitsAndBytes library for 4-bit quantisation
  • Implements left side padding and EOS/BOS tokens in tokenisation
  • Supports parallel processing on multiple GPUs
  • Integrates LoRa from the Peft library, targeting all linear layers
  • Implements gradient checkpointing
  • Uses WandB for logging

prep_and_prompt.py

• Converts text files to JSONL format
• Features:
  • Segments text files and tokenises the segments
  • Generates prompts using LLaMA2-13B
  • Implements API polling for result retrieval
  • Utilises RunPod endpoint for LLaMA2-13B
  • Uses two ThreadPoolExecutors with 30 workers each for parallel processing
  • Processes both executors simultaneously
  • Outputs generated JSON segments in JSONL format

collect_data.py

• Collects and preprocesses data, outputs as text file

clean_jsonl.py

• Cleans errors in output files from prep_and_prompt.py

inference.py

• Script to run the fine-tuned model

upload_model_hf.py

• Uploads the model to HuggingFace