CharlieLLL/paper-baselines-20260920

Paper baseline evaluation archive

Public archive of 108 selected configurations / 20,808 graded outcomes on BrowseComp-Plus150, DR-9K256, MuSiQue300 and FRAMES150. It includes complete run traces, raw predictions, grading correspondence, frozen retrieval assets and the evaluated runtime image. These are fixed-corpus local evaluation slices, not official full online benchmark results.

Code and reproduction guide: ys-2020/miles, paper/baselines. Frozen source commit: db33363ab57b. The final Git release manifest pins this dataset's immutable commit and contains the remote integrity receipt.

• Accuracy, including every control and routing operating point
• Large/small model token counts, all records, role-level CSV
• Method definitions, reproduction steps, token/cache accounting
• Run/question/trace correspondence, experiment and repair history
• Archive catalog, all-file manifest, model and asset pins

Contents

There are 167 gzip tar archives: 132 run archives, 17 failed/superseded attempt archives, 17 retrieval archives, and one reproduction bundle. Exactly 108 run archives are selected table cells. An optional SHA256-pinned runtime image is stored separately. The archive, retrieval and image payload is about 77 GiB; individual cells can be fetched separately. Model weights are pinned to their own model repositories and are not duplicated here.

episode-index.jsonl.gz joins (cell_id, full_question_sha256) to the raw result line, grading sheet/verdict and each trace member/archive. file-manifest.jsonl.gz and per-archive .manifest.json receipts record member SHA256/byte counts. A compact copy of the episode index is also in Git. Legacy truncated question previews are not join keys.

The study includes Full OARS, native MATPO, RouteLLM with a public Qwen3-8B/MiniMax-M2.7 pair under MATPO worker routing and AFM per-generation routing, plus explicitly labeled OARS/AFM controls. Harnesses, model allocation, routing units and budgets differ; keep the method labels with the results. TRAIN calibration, smoke tasks and failed attempts are separate from selected evaluation outcomes.

Token records preserve input/output by model and role, measured server cache counters and separate trace-derived ideal cache estimates. There is no measured API invoice for these locally served models. Monetary estimates require explicit model prices and cache-read rates.

The AFM DR-9K selected recovery retains 253 healthy episodes and resumes three interrupted retrieval episodes from their exact prior prefixes. Its trace dependencies span original and recovery archives and are resolved by the episode index/downloader.

Download

Evaluation artifacts are public; no login is needed. Use baselines/scripts/fetch_artifacts.py --anonymous from the linked Git branch. The downloader pins the release revision, checks archive SHA256 values and safely extracts them. The original model repositories may have separate access requirements. See the reproduction guide for CPU verification and fresh Slurm evaluation.

The owner explicitly authorized public publication of this new dataset after the initial private upload reached its storage quota. Upstream code, data, model and image terms continue to apply; this archive does not relicense third-party assets.