LLM Lab

A repo for building, fine-tuning, and deploying custom LLMs for domain-specific tasks.

🧪 Projects

Project	Description	Status
fuelrx-llm	Meal plan generation for FuelRx	🚧 In Progress

🏗️ Repository Structure

llm-lab/
├── README.md                # This file
├── fuelrx-llm/              # Meal planning LLM project
│   ├── notebooks/           # Exploration & learning
│   ├── data/                # Training data pipeline
│   ├── training/            # Fine-tuning scripts
│   ├── inference/           # Production inference
│   └── evaluation/          # Benchmarking suite
├── [future-project]/        # Next domain-specific LLM
└── shared/                  # (future) Common utilities

🎯 Philosophy

Each project in this lab follows a similar pattern:

1. Explore — Understand base model capabilities via notebooks
2. Extract — Build training data from existing systems
3. Evaluate — Create domain-specific benchmarks
4. Fine-tune — Train with LoRA for efficiency
5. Deploy — Ship to HuggingFace Endpoints

🚀 Getting Started

Each project is self-contained with its own dependencies and setup instructions. Navigate to a project directory and follow its README:

cd fuelrx-llm
cat README.md

1. Copy the example and fill in your values:

cp fuelrx-llm/.env.example fuelrx-llm/.env

2. Rebuild the Docker image:

cd fuelrx-llm

docker build -t fuelrx-llm-dev -f docker/Dockerfile.dev .

3. Start container with bash shell:

docker run -it \
  -v $(pwd):/workspace \
  --env-file .env \
  fuelrx-llm-dev /bin/bash

# ---------------------------------
# OR if you already have the container running, you can exec into it:
# Find container ID
docker ps

# Exec into it
docker exec -it <container_id> /bin/bash

# ---------------------------------
# OR start the container without immediate access to its terminal to just play with jupyter notebooks:
docker run -it \
  -v $(pwd):/workspace \
  -p 8888:8888 \
  fuelrx-llm-dev

Extracting llm logs from FuelRx

Inside the running container:

cd data
python extract_from_llm_logs.py --output training_data/meals.jsonl

To exit out and stop the container:

exit

# List running containers
docker ps

# Stop a container by name or ID
docker stop <container_name_or_id>

📦 Shared Components (Planned)

As patterns emerge across projects, common code will be extracted to shared/:

• Evaluation harnesses
• Training utilities
• Deployment scripts
• Schema validation helpers

📝 Adding a New Project

1. Create a new directory: mkdir my-new-llm
2. Copy the structure from an existing project as a starting point
3. Update this README's project table
4. Customize for your domain

🔗 Resources

• HuggingFace Transformers
• Outlines (Structured Generation)
• PEFT (LoRA)
• vLLM (Fast Inference)