SohniSwatantra / autoresearch-local-llm

What is autoresearch-local-llm?

This Python project ports Andrej Karpathy's autoresearch experiment—where an LLM iteratively tweaks a GPT training script, runs quick 5-minute experiments, and keeps only improvements in validation bits-per-byte—to run entirely locally with a free LLM like Qwen 3.5 via Ollama. No cloud APIs or Claude costs: it shares a single GPU between the researcher LLM and training, dialing back model size for feasibility on 24GB+ VRAM hardware. Fire it up locally after pulling the Ollama model or deploy to Nosana's decentralized GPUs with a one-click job config.

Why is it gaining traction?

It slashes costs to zero versus Karpathy's original API-dependent setup, letting you loop experiments indefinitely on your own hardware or cheap Nosana rentals—think $8 for 100 runs. Developers dig the full autonomy: git-commits changes, auto-resets failures, and logs progress in a TSV, echoing vibes from Karpathy's zero to hero course and nanoGPT GitHub repo. Easy swap of Ollama models keeps it flexible without vendor lock-in.

Who should use this?

ML hobbyists or researchers prototyping GPT architectures on personal NVIDIA/Apple Silicon GPUs, inspired by Karpathy's LLM council ideas or micrograd/nanogpt repos. Teams short on cloud budgets experimenting with depth, batch sizes, or attention patterns in 5-minute bursts. Fans following his YouTube, Twitter/X, or blog for hands-on autoresearch without $0.05+ per run.

Verdict

Early days at 30 stars and 1.0% credibility score—docs are solid but lacks tests and broad validation—but a clever, cost-free spin on Karpathy's framework worth forking if you've got the VRAM. Tinkerers: clone and iterate; others, wait for more polish.

(198 words)