autoresearch-stack 0.7.3

Autonomous LLM training research stack - build and ship autonomous research systems

Autonomous Research Stack

Build and ship autonomous LLM training research systems

Version: v0.7.3 | License: MIT | Python: 3.11+

An autonomous research stack for continuously improving LLM training through automated experimentation. Inspired by Karpathy's autoresearch, designed for single-GPU research labs.

---

Quick Start

# Install
pip install autoresearch-stack

# Or from source
git clone https://github.com/iknowkungfubar/autoresearch-stack.git
cd autoresearch-stack
pip install -e .

# Configure (at least one API key)
export ANTHROPIC_API_KEY=sk-ant-...
# or: export OPENAI_API_KEY=sk-...

# Run the data pipeline
autoresearch --prepare-only

# Run 10 autonomous experiments
autoresearch --experiments 10

# Run with custom config
autoresearch -c my_config.yaml -i training_data.txt --experiments 100

# Python module syntax also works
python -m autoresearch --help

Demo: Numpy Training (no GPU required)

# Test the training pipeline without PyTorch
python train_any_llm.py --demo

This runs a complete training loop using the numpy demo model, exercising the curriculum scheduler, loss tracking, and convergence detection — no GPU or PyTorch needed.

Features

Data Pipeline

Module	What it does
data_intelligence.py	Corpus cleaning, noise detection, text repair
synthetic_data.py	LLM-powered generation with Evol-Instruct
curriculum.py	Adaptive scheduling (linear, exponential, step, adaptive)
storage.py	SQLite experiment database with JSONL fallback

Experiment Engine

Module	What it does
memory.py	Vector store with semantic search (ChromaDB optional)
prioritization.py	Bandit-based selection (UCB1, epsilon-greedy, Thompson)
hypothesis.py	LLM-driven hypothesis generation with rule-based fallback
feedback.py	Reward computation, failure classification (13 types)
multi_agent.py	Multi-agent architecture (research, hypothesis, execution, evaluation)

Infrastructure

Module	What it does
sandbox.py	Safe code execution with AST-based validation
checkpoint.py	State persistence and resume
monitor.py	Real-time status and progress bars
daemon.py	Background execution with health checks and auto-restart
distribute.py	Multi-node cluster management (Docker/K8s)

LLM Integration

Module	What it does
providers.py	17+ LLM providers (Anthropic, OpenAI, OpenRouter, Ollama, vLLM, etc.)
orchestrators.py	7 agent orchestrators (CrewAI, AutoGen, LangChain, etc.)
train_any_llm.py	Training abstraction (numpy demo + optional PyTorch)

Reporting & Analysis

Module	What it does
report.py	Markdown experiment reports with comparison
figures.py	Matplotlib visualizations with graceful fallback
stats.py	Summary statistics and convergence analysis
paper.py	Research paper generation (Markdown/LaTeX)
peer_review.py	Peer review simulation (5 reviewer profiles)

---

Configuration

All configuration lives in config.yaml. Environment variables override YAML values:

export ANTHROPIC_API_KEY=sk-ant-...   # API key (never put in config file!)
export EXPERIMENT_BUDGET=1000          # Override max experiments
export LEARNING_RATE=0.0005            # Override model LR
export SYNTHETIC_USE_LLM=true          # Enable LLM data generation
export MEMORY_ENABLED=true             # Enable vector memory

Provider Support

Cloud: Anthropic (Claude), OpenAI (GPT-4/4o), OpenRouter, Google Vertex AI, Azure OpenAI, Mistral AI, Cohere, Zen AI

Local: Ollama, vLLM, LM Studio, llama.cpp, LiteLLM, KoboldCPP, LocalAI, Text Generation WebUI

Orchestrators: OpenCode, OpenCrew, AgentForge, CrewAI, AutoGen, LangChain, LlamaIndex

The Metric

val_bpb (validation bits per byte) — Lower is better. The single optimization target.

Project Constraints (Never Changed)

1. val_bpb is the ONLY metric
2. ONE change per experiment
3. Revert on regression
4. Single-GPU focused

Development Status

|| Version | Status | Tests | Coverage | Type Safety | ||---------|--------|-------|----------|-------------| || v0.7.3 | Current | 148 ✅ | 73% | 0 mypy errors | | v0.7.2 | Shipped | 104 ✅ | 57% | 43 errors | | v0.7.0 | Shipped | 53 ✅ | — | — |

Testing

# Run all tests
pytest tests/ -q

# With coverage
pytest tests/ -q --cov=./

# Run specific test file
pytest tests/test_providers.py -v

Docker

docker build -t autoresearch-stack .
docker run --rm -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY autoresearch-stack

# Multi-node cluster
docker compose up

References

• Karpathy autoresearch — val_bpb metric
• Ouroboros — Self-modifying systems
• AI Scientist — Paper generation

---

Contributing

Contributions are welcome! Please read CONTRIBUTING.md for detailed guidelines on our development process, coding standards, PR workflow, and code of conduct.

License

MIT — see LICENSE for details.