hmnshudhmn24/allm-assistant

AllM Assistant

AllM Assistant is a lightweight instruction-tuned LLM focused on fitness and lifestyle guidance. This repository contains training scripts, inference wrappers, a Gradio demo for Hugging Face Spaces, and a sample instruction-response dataset to fine-tune a causal LM (GPT-2 by default).

Contents

• src/ — model/inference/training utilities
• data/ — sample train.jsonl and val.jsonl
• hf_space/ — Gradio demo app
• requirements.txt — exact package versions to reproduce the environment
• README.md, model_card.md, LICENSE, .gitignore

Quick start (local)

1. Create and activate a virtual env:

   python -m venv venv
   source venv/bin/activate   # Windows: venv\Scripts\activate
   pip install -r requirements.txt
   

2. Train (small quick demo):

   python src/trainer.py --model_name_or_path gpt2 --train_file data/train.jsonl --validation_file data/val.jsonl --output_dir outputs/allm --num_train_epochs 1 --per_device_train_batch_size 1
   

3. Inference:

   python src/inference.py --model_dir outputs/allm --prompt "Create a 10-minute beginner home workout for fat loss."
   

4. Run the demo locally:

   python hf_space/app.py
   

Notes

• This project uses GPT-2 by default for speed. After testing, you can replace the base model with larger OSS LLMs.
• For efficient fine-tuning on limited hardware, consider using PEFT/LoRA (PEFT is included in requirements).
• The dataset included is synthetic sample data for demo and testing only — expand with high-quality real data for production.