# Reconstructing problem text This release contains **no upstream problem text, gold answers, reference solutions, or answer options** (see `docs/license_audit.md`). Every table keys on stable identifiers instead. This document explains how to attach the text yourself from the upstream sources, under those sources' own licenses. ## The two identifier columns Every per-problem table carries both: - `problem_id` — the canonical, cross-solver identifier. Use this to join the same problem across the gpt-oss-120b and Gemma-4-31B-it settings. - `source_problem_id` — the identifier exactly as it appeared in that setting's own run artifacts. **This is what you match against upstream.** They differ because the two solver campaigns were exported under different naming schemes. For OmniMath and LAB-Bench the two solvers used *different* `source_problem_id` namespaces for the very same problems: | Benchmark | gpt-oss-120b `source_problem_id` | Gemma-4-31B-it `source_problem_id` | canonical `problem_id` | |---|---|---|---| | omnimath2 | `omni2_1` | `omni2:t01:1` | `omni2_...` | | labbench | `labbench_cloningscenarios_000001` | `lab-bench:CloningScenarios:00540e26-...` | `labbench_...` | | jeebench | `jeebench:JEE Adv 2016 Paper 1:1` | same | same | | scibench | `scibench:atkins:e1.1(a)(a)` | same | same | Both solvers cover an identical canonical problem set in every paired setting (verified: symmetric difference is empty for all five pairs). Always join on `problem_id`, never on `source_problem_id`, when comparing solvers. `data/problems.csv` carries both columns for all 6,803 released problems, plus a `legacy_tier_id` column for OmniMath (`omni2:tNN:idx`) that matches the Gemma namespace and the primary-split identifiers. ## Per-benchmark reconstruction ### Omni-MATH-2 (4,181 problems) Upstream: (Apache-2.0). The released slice is a filtered exact-answer subset built locally; it is not a contiguous upstream range, so you cannot slice it by index. Use `data/problems.csv` (`source`, `difficulty`, `difficulty_tier`, `domain`) to identify rows, then match against the upstream corpus. The `legacy_tier_id` column encodes the project's own tier partition: `omni2:tNN:idx` is the `idx`-th problem of difficulty tier `NN`. Tier sizes are in `registry/release_manifest.json` under `coverage.omnimath_tier_counts`. Caveat: reconstructing this slice exactly requires the project's filtering script, which is not part of this dataset release. It is in the code repository at . ### JEEBench (515 problems) Upstream: (MIT). `source_problem_id` has the form `jeebench::`, e.g. `jeebench:JEE Adv 2016 Paper 1:1`. Split on `:` to recover the upstream paper name and 1-based question index. All 515 public problems are covered. ### SciBench (565 problems) Upstream: (MIT). `source_problem_id` has the form `scibench::`, e.g. `scibench:atkins:e1.1(a)(a)`. The textbook segment names the JSON file under `dataset/original/`. The project's text-only slice holds 574 rows carrying 571 distinct identifiers (three identifiers are reused, see below); 565 of those 571 were executed under all four protocols and are released. The six unexecuted identifiers are all from the `thermo` textbook: `scibench:thermo:1.3`, `1.6`, `2.13`, `6.10`, `9.9`, and `14.5`. Why those six were not run is not recorded in the artifacts this release was built from. Three SciBench source identifiers are ambiguous upstream — `scibench:quan:2.13`, `scibench:stat:5.8-5`, and `scibench:stat:Problem 1.1.1` each name two distinct problems. The canonical `problem_id` disambiguates them with a content-hash suffix (`...#<12 hex>`); the plain `source_problem_id` does not. When reconstructing those three, expect two upstream candidates and disambiguate by content. ### LAB-Bench (1,542 problems, in two nested slices) Upstream: (CC-BY-SA-4.0), revision `5c77cec648430f30611808808861eb86f81d5eaa`, `train` split. Two slices are released, and the smaller is a strict subset of the larger: - `llm_strict` (741): subsets `CloningScenarios` (33), `ProtocolQA` (108), `SeqQA` (600). - `text_no_tool` (1,542): the strict slice plus `DbQA` (520), `LitQA2` (199), `SuppQA` (82). `FigQA` (181) and `TableQA` (244) are excluded upstream-side because they require image assets. There are no row-level exclusions within the included subsets. The Gemma-namespace `source_problem_id` embeds the upstream UUID directly: `lab-bench::`. The gpt-oss namespace (`labbench__`) does not; use `problem_id` and `data/problems.csv` (`subset` column) to bridge. **Answer-option order.** The runs used a deterministically shuffled option order derived from the upstream row id and ideal answer, and the gold label in the runs is the *shuffled* letter. If you reconstruct from upstream you will get the upstream option order, and your letter labels will not match. This release does not contain the shuffle, so per-option correctness is not reconstructible from this release alone. The per-problem correctness flags here are unaffected — they record whether the protocol's chosen option was the correct one. **Do-not-train request.** LAB-Bench ships a canary and an upstream request not to train on it. This release contains no LAB-Bench content, only identifiers. Do not use these identifiers to assemble a training corpus that would violate that request. ## What you cannot reconstruct from this release - **Raw model generations.** No protocol produced text is included. Per-protocol correctness is the finest-grained outcome released. - **Baseline final answers.** The confidence probe's prompt included the baseline's final answer as untrusted input. That answer text is not in this release, so `data/probe_inputs.jsonl` carries `baseline_final_answer: null`. The probe is therefore **not re-runnable** from this release alone; see `docs/schema.md`. - **Per-protocol token/call/wall-time costs for the nine non-primary settings.** Cost accounting is released only through the aggregate tables. The primary OmniMath 423-problem held-out analysis in `data/aggregate/ main_routing_heldout.csv` reports costs, but the per-problem cost columns behind it are not staged here.