Datasets:

tomyimkc
/

repro-optimal-regret-for-policy-optimization-in-contextual-bandits-traces

Modalities:

Formats:

Size:

Tags:

Libraries:

Dataset card Data Studio Files Files and versions

Claude Code traces preview

Sat, Jul 25

Reproduce one ICML 2026 paper for the Open Reproductions challenge. Your paper directory: ~/icml26-factory/papers/CXsD6Ygc8e Read ~/icml26-factory/papers/CXsD6Ygc8e/paper.json FIRST — it has the title, arxiv id, area, abstract, and the OFFICIAL CLAIM LIST you must target (field "claims", an array of claim strings, in order). Read the full brief at ~/icml26-factory/AGENT_BRIEF.md and follow it exactly. It defines scoring, the honesty rules, the experiment-script contract, and the results.json schema. Non-negotiables: - ONE experiment script per official claim: experiments/exp_claim<N>.py, matching claim order. - Only numpy/scipy/pandas/stdlib. No torch, no network at runtime, no model downloads, no GPU. - Seed everything. Deterministic output required. - Do NOT put elapsed/wall-clock times in RESULT_JSON — they are not reproducible. - Each script should finish in about 1-3 minutes. Prefer vectorised numpy at large n over long Python loops: scale is what separates a 2-point 'verified' from a 1-point 'toy'. - Include a control that relaxes the claim's precondition and show it degrades. If the control does NOT degrade, say so honestly and consider whether the verdict should be 'toy' rather than 'verified' — a control that never breaks means the mechanism was not isolated. - Every official claim needs its own experiment. An untargeted claim scores ZERO. - PROSE NUMBER RULE: every decimal figure you write in method/finding/posterFinding/scopeNote/executiveSummary/conclusion must trace to your script's RESULT_JSON, to the official claim text, or to a constant in your script source. If you want to discuss a number, emit it in RESULT_JSON first. The builder enforces this and will fail your build. Use ~/.venvs/icml26/bin/python for everything. Read the paper: curl https://arxiv.org/html/<arxiv>, or /abs/, or the export API. If the full text is unavailable, work from the official claims plus the abstract and say so in your scope notes. Validate with: ~/.venvs/icml26/bin/python ~/icml26-factory/lib/build_logbook.py ~/icml26-factory/papers/CXsD6Ygc8e Iterate until it exits 0. Do NOT pass --publish. Honesty over score: an inflated verdict will be overturned by the independent judge and costs us credibility. A truthful 'toy' beats a false 'verified'. If a claim cannot be targeted on CPU, mark it inconclusive and explain why. 2 messages 73 tools

Agent traces

Agent sessions published from a Trackio Logbook.

Downloads last month: 114

Number of rows:

Total file size: