autoresearch-nano 1.0.0

nanoGPT-based autonomous ML research loop (research prototype).

autoresearch — nanoGPT-based autonomous ML research loop (research prototype)

What this is: a fork of Karpathy's nanoGPT / nanochat plus an autonomous research-orchestration layer driven by a Claude Code program.md playbook. The goal is to have an LLM agent iterate on train.py inside a fixed time budget and hunt for lower val_bpb.

What this is NOT (v1.0.0): a finished Tribunal-evaluated, Silicon-fused research stack. Read the honest-scope section before depending on any specific feature.

---

Honest scope for v1.0.0

This release ships a clean installable skeleton around the parts that really work, plus honest documentation of what does not.

Component	Status	Notes
nanoGPT-style training core (train.py)	Works	Requires H100-class GPU and flash-attn3 kernels for real runs.
CPU-friendly toy training (autoresearch toy-train)	Works	No GPU required. Useful for CI and smoke-tests.
CLI (info, config, toy-train)	Works	No network, no GPU.
JSON metrics sidecar (run_metrics.json)	Works	Replaces brittle stdout-regex parsing.
P2PCLAW Silicon integration	PARTIAL (inert)	silicon/grid_generator.py is a placeholder. Do not rely on it.
Tribunal evaluation	NOT IMPLEMENTED	Stub only (autoresearch.tribunal), raises on use.
Railway P2PCLAW API	OPTIONAL	Core loop and tests do not need it.

See ROADMAP.md for the full gap list.

Install

pip install autoresearch
# or, with optional extras:
pip install "autoresearch[full]"   # transformers, datasets, wandb
pip install "autoresearch[dev]"    # pytest, build, twine

Requires Python >= 3.10. Core deps: torch, numpy, requests.

Quickstart (CPU, no GPU, no network)

# 1. See environment + dependency status
autoresearch info

# 2. Show the default config
autoresearch config

# 3. Run a tiny nanoGPT training smoke-test (writes run_metrics.json)
autoresearch toy-train --steps 20 --verbose

You should see the loss drop across steps and a run_metrics.json written to the current directory.

Running the real nanoGPT training loop

The top-level train.py is unchanged from the nanochat-derived original and requires an NVIDIA GPU with flash-attn3 kernels (H100 recommended; other Hopper/Blackwell-class cards may work via kernels-community/flash-attn3). CPU is not supported for train.py — use autoresearch toy-train for CPU.

uv run prepare.py      # prepares data under ~/.cache/autoresearch/
uv run train.py        # runs the 5-minute time-budgeted training loop

The autonomous research loop itself is orchestrated via Claude Code following program.md.

Structured metrics (preferred over stdout regex)

Earlier versions parsed training results out of stdout using regex against the --- block. That path is now a fallback. Primary path:

from autoresearch.metrics import write_metrics, read_metrics

write_metrics("run_metrics.json", {"val_bpb": 0.9979, "num_steps": 953})
read_metrics("run_metrics.json")  # -> {"val_bpb": 0.9979, "num_steps": 953}

The legacy parser is still available as autoresearch.metrics.parse_stdout_summary(text) for old run logs.

Tribunal and Silicon

• autoresearch.tribunal.evaluate(...) raises TribunalNotImplementedError.
• silicon/grid_generator.py does not currently emit a usable grid.

Both are tracked in ROADMAP.md.

Testing

pip install -e .[dev]
pytest

All tests are offline and CPU-only. CUDA-requiring cases are skip-marked.

License

Apache License 2.0 — see LICENSE.

Credits

• Andrej Karpathy — nanoGPT and nanochat (MIT).
• Francisco Angulo de Lafuente (Agnuxo1) — autoresearch orchestration layer, P2PCLAW integration work.