Koji Ono

You can run the full LLM research lifecycle: build method-by-dataset literature matrices, scaffold PyTorch/Hydra experiment projects, execute seeded ablation runs with reproducibility hashes, probe attention via TransformerLens, draft NeurIPS/ACL papers, and publish bundles to Zenodo or HuggingFace with DOIs.

LLM researchers and ML engineers working on evaluation benchmarks, fine-tuning, prompt/ICL studies, or attention and architecture internals. It suits single-GPU tinkerers wanting overnight autonomous exploration as well as teams comparing multiple project versions with statistical rigor.

Tinker mode iteratively edits a single train.py under a fixed wall-clock budget (default 5 minutes per cycle) optimizing val_bpb on one GPU. Swarm mode runs N parallel agents (default 3) with distinct exploration strategies, aggregating results into a shared global best via an orchestration script.

Experiments scaffold on uv-based Python projects with PyTorch, HuggingFace transformers/datasets/accelerate, and Hydra. Interpretability requires TransformerLens or nnsight. Publishing needs HF_TOKEN or ZENODO_ACCESS_TOKEN environment variables. Visualization builds on MkDocs Material v0.17.0+ with Chart.js.

Each run's USD cost is computed from duration times GPU unit price against an editable price table, with warnings at 80% budget and rollback proposals at overrun. Reproducibility is enforced via run_id, git revision, config hash, uv.lock snapshots, and preserved failed-run postmortems.