Implementing QLoRA and best-model checkpointing (2 issues)

[dcrey7]

🚀 Key Feature 1: QLoRA for Efficient Fine-Tuning

• 4-bit Quantization: The training script now loads the base model in 4-bit precision using the bitsandbytes library.
• Low-Rank Adaptation (LoRA): A PEFT adapter is applied to the quantized model. This freezes the 4-billion base parameters and trains only a tiny fraction of new parameters (~0.47%).
• Benefit: This dramatically reduces the VRAM requirement from >16GB to ~5-6GB, making it possible to fine-tune this model on a wide range of consumer and prosumer GPUs.

💾 Key Feature 2: Best-Model Checkpointing

• Validation Loop: The script now runs a validation loop at the end of each epoch to calculate the validation loss.
• Save on Improvement: A checkpoint of the trained LoRA adapter is saved to disk only when the validation loss is lower than the previous best.
• Benefit: This ensures that the final saved artifact is the best-performing version of the model from the entire training run, not just the last one.

📈 Other Improvements:

• Robust Data Loading: The get_dataloaders function has been improved with more robust error handling.
• Cleaned .gitignore: The wandb/ directory has been added to .gitignore to prevent local log files from being committed.
• Enhanced Code Comments: The code in config.py and train.py has been commented to make it easier for future users to switch between test runs and full training sessions.

was able to run it on google colab - https://colab.research.google.com/drive/1H4cLS-HotsaS_40Bfo_7d_fZVwqBcrLW#scrollTo=yiE6lvpvdmio, i was able to get save the final lora adapters in the output file with best models, will hape to add your own hugging face token here

These changes make the project significantly more practical and user-friendly, especially for those without access to high-end industrial hardware. looking forward to my first commit , feedback is really appreaciated will try my best