llm-forge

Build your own AI model. No ML expertise required.

llm-forge is a conversational AI assistant that helps you build custom language models. Just type llm-forge and tell it what you want — it handles everything: data cleaning, training, evaluation, and deployment.

Install (One Command)

curl -fsSL https://raw.githubusercontent.com/Nagavenkatasai7/llm-forge/main/install.sh | bash

Then open a new terminal (or run source ~/.zshrc) and type:

llm-forge

That's it. LLM Forge will guide you through building your AI model.

Already have Python 3.10+? You can also install directly:

pip install llm-forge-new
llm-forge

What Happens When You Run It

$ llm-forge

╭─ LLM Forge — Build your own AI model ─╮
│  Just tell me what you want to build.  │
╰────────────────────────────────────────╯

You: I want to build a customer support chatbot

Forge: I'll set that up. Let me check your hardware...
       Found: NVIDIA RTX 4090 (24 GB VRAM)

       I recommend Llama-3.2-1B with LoRA fine-tuning.
       Where is your training data?

You: ~/data/support_tickets.jsonl

Forge: Found 2,400 records in Alpaca format.
       Config saved. Starting training...
       ████████████░░░░░░  Step 142/300  loss: 1.42  ETA: 8 min

LLM Forge works as a conversational manager — you talk to it, it takes action. No YAML editing, no CLI commands to memorize.

No API key? LLM Forge includes a free guided wizard that works offline — no API needed.

Want the full AI-powered experience? Set a Claude API key:

export ANTHROPIC_API_KEY=your-key-here
llm-forge

Quick Demo (5 Minutes, No Config)

llm-forge demo

Trains a tiny 135M model on bundled sample data. Works on CPU. Zero configuration.

Prerequisites

• Python 3.10+ (the installer handles this)
• NVIDIA GPU with 8+ GB VRAM (recommended) or CPU for testing
• 20 GB free disk space

No GPU? You can test on CPU with small models, or use Google Colab for free GPU access.

Installation Options

# Option 1: One-line installer (recommended — handles everything)
curl -fsSL https://raw.githubusercontent.com/Nagavenkatasai7/llm-forge/main/install.sh | bash

# Option 2: pip install (if you have Python 3.10+)
pip install llm-forge-new

# Option 3: From source (for development)
git clone https://github.com/Nagavenkatasai7/llm-forge.git
cd llm-forge
pip install -e ".[all]"

How It Works

You edit this:                  llm-forge handles the rest:
+-----------------+
|   config.yaml   |  ------>   Data Cleaning -> Training -> Evaluation -> Serving
+-----------------+

1. Generate a starter config:

llm-forge init --template lora

2. Edit config.yaml - set your model, data path, and hyperparameters. Every field has inline documentation.

3. Validate before training (catches errors early):

llm-forge validate config.yaml

4. Train:

llm-forge train --config config.yaml

llm-forge auto-detects your hardware and optimizes accordingly.

Example: Train on Your Own Data

Create a JSONL file with instruction-output pairs:

{"instruction": "Summarize this article", "input": "The economy grew...", "output": "Economic growth..."}
{"instruction": "Translate to French", "input": "Hello world", "output": "Bonjour le monde"}

Then create a config (config.yaml):

model:
  name: "meta-llama/Llama-3.2-1B"
  max_seq_length: 2048

data:
  train_path: "./data/train.jsonl"    # Your data file
  format: "alpaca"

training:
  mode: "lora"                        # Memory-efficient fine-tuning
  output_dir: "./outputs/my-model"
  num_epochs: 3
  learning_rate: 2.0e-4

Run llm-forge train --config config.yaml and you're done.

Features

Getting Started

Training

Data & Evaluation

Production

Recommended Starting Models

Hardware Compatibility

CLI Reference

llm-forge init [--template lora|qlora|pretrain|rag|full]   # Generate starter config
llm-forge validate config.yaml                              # Validate + hardware check
llm-forge train --config config.yaml [--verbose] [--dry-run]  # Train model
llm-forge eval --config config.yaml                         # Run benchmarks
llm-forge serve --config config.yaml                        # Launch chat UI
llm-forge export --config config.yaml --format gguf         # Export model
llm-forge clean --config config.yaml                        # Data cleaning only
llm-forge rag build --config config.yaml                    # Build RAG index
llm-forge rag query "question" --config config.yaml         # Query RAG
llm-forge info                                              # System + GPU info
llm-forge hardware                                          # Hardware summary

Example Configs

Ready-to-use configs in configs/:

Troubleshooting

Docker

# GPU training
docker build -t llm-forge:gpu -f docker/Dockerfile.gpu .
docker run --gpus all -v $(pwd)/outputs:/app/outputs \
  llm-forge:gpu train --config configs/demo_lora_llama.yaml

# Docker Compose (training + serving + vector DB)
docker compose -f docker/docker-compose.yml up

HPC / SLURM

# Single GPU on Hopper cluster
sbatch scripts/slurm/train_demo.sbatch

# Multi-node distributed
sbatch scripts/slurm/train_multi_node.sbatch

See Hopper Cluster Guide for detailed setup.

Documentation

Contributing

See CONTRIBUTING.md for development setup and guidelines.

Citation

@software{llm_forge,
  author = {Chennu, Naga Venkata Sai},
  title = {llm-forge: Config-Driven LLM Training Platform},
  year = {2026},
  url = {https://github.com/Nagavenkatasai7/llm-forge}
}

License

Apache License 2.0

Author

Naga Venkata Sai Chennu - George Mason University

• GitHub: @Nagavenkatasai7
• Email: nchennu@gmu.edu