🧠 Training-Model

A hands-on repository to experiment with training/fine-tuning large language models like Mistral-7B using Hugging Face, LoRA, and custom datasets.

📂 Project Scope

• Fine-tuning transformer models
• Hyperparameter optimization
• Using datasets from local and custom inputs
• Learning rate scheduling and precision handling
• Experiment tracking and analysis (you can wire in W&B)

🚀 Tech Stack

• Python
• Transformers (Hugging Face)
• Datasets (JSONL)
• PEFT / LoRA
• 4-bit k-bit training (optional)

✅ Quickstart

1. Create an environment and install deps

make install

2. Run a quick smoke test (no heavy deps needed)

make smoke

3. Fine-tune with LoRA (edit the YAML first!)

make train

• Config: configs/mistral_lora_example.yaml
• Output: outputs/mistral7b-lora-demo

🗂️ Data Format

Provide a JSONL file where each line contains prompt and response fields. See data/sample.jsonl:

{"prompt": "What is LoRA?", "response": "Low-Rank Adapters..."}

The trainer will render each pair into this template by default:

<s>[INST] {prompt} [/INST] {response}</s>

You can change the template via text_template in the config.

⚙️ Config Highlights

See configs/mistral_lora_example.yaml:

• base_model_name_or_path: e.g. mistralai/Mistral-7B-Instruct-v0.2
• use_4bit: toggle 4-bit k-bit training (requires CUDA + bitsandbytes)
• LoRA: lora_r, lora_alpha, lora_dropout, target_modules
• Training: batch size, lr, epochs, scheduler, precision

🧪 Scripts

• src/training/train_lora.py: LoRA fine-tuning script
• src/data/build_dataset.py: JSONL utilities and HF dataset builder
• tests/smoke_test.py: minimal data utility test
• scripts/clone_cashew.sh: optional utility to clone a related HF Space

📌 TODO

• Add cleaned dataset
• Upload training logs / metrics
• Automate preprocessing scripts
• Experiment with smaller models (e.g. Falcon-1B)

---

Made with ☕, curiosity, and experimentation.