ofou's picture
|
download
raw
788 Bytes

Neural Thickets tinygrad reproduction

Pure tinygrad port of the official RandOpt 1D suite (simple_1D_signals_expts) for ICML 2026 Agent Repro (paper arXiv 2603.12228 / OpenReview 92oF5bU4cU).

Setup

cd tinygrad_repro
uv venv .venv --python 3.12
uv pip install --python .venv/bin/python tinygrad numpy trackio huggingface_hub

Runs

.venv/bin/python run_claim1_3.py   # Claims 1 & 3 (needle/thicket/plateau)
.venv/bin/python run_claim2.py     # Claim 2 width/depth proxy
.venv/bin/python run_claim4_5.py   # Claims 4 & 5 (RandOpt vs ES/SGD + population)

Results land in outputs/claim*/results.json.

Notes

  • HF Jobs remain 402 (insufficient credits) — no Qwen2.5 / PPO / GRPO GPU runs.
  • Metrics use AR rollout MSE (matches official eval).

Xet Storage Details

Size:
788 Bytes
·
Xet hash:
f4e35e5249f5381b2ee0e573a86ed84839290dfa7c9a01ed7f7edbd73c0e5d14

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.