David-beakr/scientific-work-execution-seed

Internal Beakr dataset — not for public distribution. Host as a private repository. See "License & attribution" before any external sharing.

Scientific Work Execution Seed v0.5.1

30 hand-curated tasks for Pillar 4: Scientific Work Execution. The pillar question: can the agent actually do scientific work — load real data files, run analyses, write code that executes, and produce verifiable outputs? This is distinct from scientific Q&A (recall about scientific topics).

A task is execution-shaped when the agent receives raw data (CSV, SDF, FASTA, …) or a code specification, must produce a derived artifact (a hypothesis, a code file, a structured-output JSON, a plot), and the artifact is graded against gold by a deterministic or hybrid check. Every task here satisfies all three.

Stats

Tasks	30
Source datasets	5 (DiscoveryBench, SciCode, ScienceAgentBench, GenoTEX, BLADE)
Domains	15
Check types	5 (one canonical type per source)
Schema version	1.0 (canonical EvalCase envelope)
License	Internal compilation; borrowed content carries its source license (see below)

By source:

Source	Tasks	Check type
ScienceAgentBench	10	program_matches_gold
DiscoveryBench	8	hypothesis_matches_gold
GenoTEX	5	gtas_pipeline_outputs_match_gold
BLADE	4	analysis_decisions_match_gold
SciCode	3	code_passes_test_cases

By domain (15): Computational Chemistry 7, Biomedical genomics 5, Meta science 3, Humanities/archaeology 2, Semiconductor physics 2, Bioinformatics 2, and one each of Biology/ecology, Sociology, Engineering, Numerical linear algebra, Psychology & Cognitive Science, Sports analytics/social science, Ecology/wildlife biology, HCI/experimental psychology, Animal behavior/anthropology.

What's in this repo

scientific-work-execution-seed/
  README.md
  gold.json    30 hand-curated tasks

No fixture corpus ships here. Each task's input data lives in its source dataset, which the harness adapter fetches separately (see "Harness requirements").

Case shape

Each task is one case in the canonical EvalCase envelope; the source-specific fields are bucketed:

{
  "id": "PILLAR4-XX-NNN",
  "input": { "...agent-visible spec: question / task_instruction / function_header / research_question, domain_knowledge, input_data_files, dataset_folder_tree, expected_outputs..." },
  "reference": { "gold_hypothesis": "...(where applicable)", "checks": [ ... ] },
  "fixture": { "...source-fetch pointers: source_id / source_trait / source_dataset / source_github_name..." },
  "fixture_private": { "gold_program_name": "...", "eval_script_name": "..." },
  "metadata": { "kind": "task", "source": "...", "domain": "...", "workflow_tags": [...], "required_libraries": [...], "curation_notes": "..." }
}

• input — what the agent sees (task spec + provided data-file names).
• reference — gold (gold_hypothesis for DiscoveryBench) plus the checks.
• fixture / fixture_private — source-fetch locators, and gold artifacts the agent must not see (gold program / eval-script names).
• metadata — provenance and harness hints (source, domain, required libraries, curation notes).
• Dataset-level sources, curation_criteria, and rejected_during_curation remain top-level keys.

The five sources

Each task borrows from one of five public benchmarks. Per source: what it is, task shape, where the gold lives, and how it was curated. Curation criteria (the explicit rules) and the rejection log follow in their own sections.

DiscoveryBench (8 tasks) — hypothesis_matches_gold

Data-driven discovery: the agent receives one or more CSV files plus a discovery question (e.g. "What factors influence the success of unintentionally introduced plants?") and must analyze the data to form a hypothesis.

• Task IDs: PILLAR4-DB-001…008.
• Task shape: {data files, domain-knowledge briefing, question, gold hypothesis}. Ingest the CSVs, run analysis (regression, time series, PCA, bootstrap), state a hypothesis matching the gold.
• Source: nhop/discoverybench (real_test reformat of allenai/discoverybench); Majumder et al., 2024, arXiv:2407.01725; ODC-BY.
• Curation: used the nhop reformat because it exposes gold_hypothesis and gold_workflow, which the original hides. Hand-curated 8 against the criteria below (rejected DB row 160 for stating exact regression coefficients).

SciCode (3 tasks) — code_passes_test_cases

Research-coding: each problem decomposes into sub-steps that each require a Python function passing test cases.

• Task IDs: PILLAR4-SC-001…003.
• Task shape: {function header, background formulas, required dependencies, test cases}. The agent writes Python that passes the deterministic asserts.
• Gold: the SciCode1/SciCode HuggingFace dataset holds the problems (headers, backgrounds, test-case expressions). The numeric target values the asserts compare against are not there — they ship as a separately-downloaded test_data.h5 in the SciCode eval repo (scicode-bench/SciCode, ./eval/data/test_data.h5).
• Source: Tian et al., 2024, arXiv:2407.13168; Apache-2.0 (NOTICE preservation required).
• Curation: picked accessible first sub-steps (plus one single-sub-step problem). Rejected Problem 77 (Berendsen-thermostat MD) and Problem 11 (GADC entanglement) as not hand-verifiable. Householder QR (SC-003) added in v0.2 to dilute semiconductor-physics over-representation.

ScienceAgentBench (10 tasks) — program_matches_gold

102 expert-curated tasks from 44 peer-reviewed publications; the agent writes a self-contained Python program that processes scientific data and produces an output artifact (CSV, PNG, …).

• Task IDs: PILLAR4-SAB-001…010. Sub-domains: Computational Chemistry 7, Bioinformatics 2, Psychology & Cognitive Science 1 (4 GIS tasks skipped — see rejection log).
• Task shape: {task instruction, domain knowledge, dataset folder tree, dataset preview, expected output filename, gold program, eval script, required libraries}. The agent's program, when executed, must produce the expected output; the eval script compares it against gold.
• Gold: gold programs and eval scripts are not in the HuggingFace dataset card — they live in the public OSU-NLP-Group/ScienceAgentBench GitHub repo (benchmark/gold_programs/, benchmark/eval_programs/), distributed as a password-protected zip. The adapter clones the repo and references gold by name (gold_program_name, eval_script_name in our gold.json).
• Source: Chen et al., ICLR 2025, arXiv:2410.05080; CC-BY-4.0.
• Curation: picked distinct comp-chem sub-fields (DeepChem drug discovery, Matminer/pymatgen materials, ProLIF molecular dynamics, pymatgen electronic structure) to avoid clustering; used the verified split.

GenoTEX (5 tasks) — gtas_pipeline_outputs_match_gold

Real-world gene-expression analysis: for a disease "trait," the agent must (1) select usable datasets, (2) preprocess them (clean, normalize, link clinical + expression data), and (3) run regression to identify significant genes. The benchmark has 1,384 problems across 132 traits; we use 5 unconditional ones (condition=None, the simplest variant), one per disease category.

• Task IDs: PILLAR4-GT-001…005 — Alzheimer's (neurological), Breast Cancer (oncology), Atherosclerosis (cardiovascular), Type 2 Diabetes (metabolic), COVID-19 (infectious/viral; recent → lower leakage risk).
• Task shape: {trait, condition=None, research question, available cohorts}. Outputs go in GenoTEX's directory structure: output/preprocess/{trait}/clinical_data/{cohort}.csv, gene_data/{cohort}.csv, linked {cohort}.csv, cohort_info.json, and output/regress/{trait}/significant_genes_condition_None.json.
• Data: the full data folder is 82 GB (Google Drive; also a Liu-Hy/GenoTEX HuggingFace mirror). Fetch only the 5 selected traits — each ~1–3 GB.
• Source: Liu et al., MLCB 2025, arXiv:2406.15341; CC-BY-4.0.
• Curation: one trait per distinct disease category; well-studied traits (reliable gold gene lists per the Open Targets Platform); multiple GEO cohorts each so dataset selection is non-trivial.

BLADE (4 tasks) — analysis_decisions_match_gold

Statistical-analysis: given a dataset + research question, the agent produces (a) conceptual variables, (b) a data-transformation function, and (c) a statistical-model function. Each component is graded against expert annotations.

• Task IDs: PILLAR4-BL-001…004 — soccer (multilevel regression; the Silberzahn et al. 2018 "red cards" study), fish (count regression, Poisson/NB), reading (mixed-effects with interaction), panda_nuts (mixed-effects with learning curve).
• Task shape: {dataset name, research question, dataset description, expected statistical methods}.
• Gold: expert annotations (transform_specs + cv_specs) are auto-prepared on first run via the blade_bench package (load_ground_truth(dataset)).
• Source: github.com/behavioral-data/BLADE; Gu et al., EMNLP 2024 Findings, arXiv:2408.09667; code Apache-2.0, data ODC-By.
• Curation: one dataset per major area (sports/social, ecology, HCI, animal behavior); visible expert ground-truth; non-trivial methodology (regression, mixed-effects, count models).

The five check types

Each task uses one canonical, hard check type.

• hypothesis_matches_gold (DiscoveryBench, 8) — hybrid. Deterministic layer: required_values (numbers/entities that must appear), alternative_phrasings, must_not_contain (a confabulation tripwire of directional significance claims — e.g. "significantly higher/lower", "strong relationship", "significant gender gap"). Judge layer: a per-task rubric stating the semantic pass conditions. The judge catches right-number-wrong-direction errors the substring layer can't. Under the scoring policy below, the must_not_contain tripwire is restricted to tokens that cannot collide with a correct answer's own surface form — notably it no longer includes the bare "significant difference", which is a substring of the legitimate null result "no significant differences" (DB-004); the null-result semantics are owned by the rubric ("A pass requires explicit statement of no significant difference").
• code_passes_test_cases (SciCode, 3) — deterministic. The agent's code plus each test_case runs in a sandbox; each asserts equality against a stored target (resolved from test_data.h5, not in our gold). All test cases must pass.
• program_matches_gold (ScienceAgentBench, 10) — three-part hard check: (1) valid execution — the program runs without errors in a sandbox with the required libraries; (2) output produced — the expected file exists at the specified path; (3) success — the source eval script's comparison against gold passes (column equality for CSVs, image similarity for PNGs, accuracy/F1 for predictions). All three must pass.
• gtas_pipeline_outputs_match_gold (GenoTEX, 5) — composite across three phases: dataset selection (cohort is_usable set-overlap ≥ 0.7 vs gold cohort_info.json), preprocessing (clinical/gene/linked CSVs match within tolerance for ≥ 80% of selected cohorts), statistical analysis (top-K significant-gene AUROC ≥ 0.6 vs gold). All three must pass. Thresholds are intentionally lenient at v0.3 — first detect whether any part runs end-to-end; tighten once integration provides baselines.
• analysis_decisions_match_gold (BLADE, 4) — composite via BLADE's evaluator: conceptual variables (semantic match), data transformation (graph match vs transform_specs), statistical model (spec match — regression family, random-effects structure — vs cv_specs). Pass requires partial success on all three. The match function is implemented in blade_bench (Apache-2.0) and uses a GPT-4o-class semantic matcher.

Hybrid scoring. Judge-bearing checks (hypothesis_matches_gold, a few SAB visualization eval scripts, and analysis_decisions_match_gold's semantic matcher) use tight per-task rubrics enumerating concrete pass conditions, and run only after the deterministic layer passes. 13 of 30 tasks (SciCode + most SAB) need no judge at all.

Scoring policy — closed-form vs. judge

Deterministic checks (exact/numeric/format/execution) gate (hard) only on closed-form properties — numbers, file outputs, exact format, code that passes asserts — where paraphrase cannot move the target (SciCode asserts, ScienceAgentBench eval scripts, GenoTEX set-overlap/AUROC). Open-ended/semantic properties (a hypothesis, a null result, an analysis decision) are owned by an LLM judge or semantic matcher against a tight, reference-guided per-task rubric, following these tasks' own source benchmarks — DiscoveryBench scores hypotheses by LLM matching and BLADE by graph + semantic matching, neither gating the judgment behind a brittle substring veto. Any lexical proxy for a semantic property is restricted to non-colliding confabulation tripwires, never a hard veto over the judge. This is grounded in the QA-evaluation literature on lexical-match false negatives, especially on negation/null results (SQuAD EM/F1 limits; Bulian et al. 2022 BEM/answer equivalence; Kamalloo et al. 2023; negation-aware evaluation, Anschütz et al. 2023), and on LLM-as-judge reliability with explicit rubrics (Zheng et al. 2023; G-Eval, Liu et al. 2023). Composition note: hypothesis_matches_gold currently requires the deterministic layer AND the judge; "judge-primary with advisory lexical signals" is a kernel capability — it would live in the spec_checks check-dispatcher proposed in the gap log (GAP-03). Until it lands, the deterministic layer is kept collision-free so it cannot veto a correct answer.

Curation criteria (per source)

The explicit rules each candidate task had to meet (also machine-readable in gold.json under curation_criteria).

• DiscoveryBench: question contains no answer values (no leakage); question doesn't name the answer entities trivially; gold_hypothesis requires real analytical work; question is unambiguous (one correct answer).
• SciCode: sub-step is the first step of a multi-step problem (or a single-sub-step problem); background gives the formula explicitly; test cases are visible and well-defined; implementation is a translation of given equations or a well-defined numerical procedure.
• ScienceAgentBench: instruction is clear, self-contained, and names the expected output file; domain is Computational Chemistry, Bioinformatics, or Psychology/Cognitive Science; required libraries are commonly available; an eval script exists with a deterministic pass/fail; domain coverage stays balanced.
• GenoTEX: trait is unconditional (condition=None); each trait is a distinct disease category; the trait has multiple GEO cohorts (selection is non-trivial); the trait is well-studied (reliable gold gene list).
• BLADE: research question is clear and unambiguous; the dataset has visible expert ground-truth (transform_specs + cv_specs); methodology is non-trivial; one pick per major area.

Rejected during curation

Source	Identifier	Reason
DiscoveryBench	row 160 (sociology)	Question stated exact regression coefficients ("decreases from 0.5024 to 0.0923") — reverse-engineerable leakage.
SciCode	Problem 77	Berendsen-thermostat MD simulation — PhD-level comp chem, not hand-verifiable.
SciCode	Problem 11	GADC entanglement / coherent information — PhD-level quantum information theory, not hand-verifiable.
ScienceAgentBench	Instances 4, 53, 54, 55	All GIS — skipped to avoid domain over-concentration with the DiscoveryBench archaeology tasks (DB-006, DB-007).

Methodology deviations from the source papers

• Scale. Each source has far more tasks than we borrow (1–10 per source) to keep verification tractable.
• SciCode sub-step extraction. We evaluate only the first sub-step of selected problems, not the intended end-to-end problem. Each sub-step is still a valid scientific-code task.
• DiscoveryBench gold editing. For DB-001, DB-002, DB-005 we lightly edited gold_hypothesis (removed a redundancy; trimmed to the asked question; dropped an Experimental-Economics breakdown the question didn't ask). Documented per task in curation_notes.
• GenoTEX unconditional-only. We use 5 unconditional problems and skip the 1,252 conditional ones (simpler analysis, easier to verify).
• BLADE pass thresholds. BLADE's evaluator produces continuous scores; we apply intentionally lenient v0.3 thresholds.
• ScienceAgentBench partial coverage. 10 of 102 tasks, weighted toward Computational Chemistry, GIS skipped.
• Hand-curation as first-class. None of the source papers treat hand-curation as part of the methodology; we treat it as load-bearing.

Harness requirements

Heavier than the other pillars. In short:

• Fetching (five mechanisms): DiscoveryBench CSVs from nhop/discoverybench (HuggingFace); SAB dataset folders + gold programs + eval scripts from the SAB GitHub repo (password-protected zip); SciCode test_data.h5 from scicode-bench/SciCode; GenoTEX per-trait data (~1–3 GB each) from Google Drive or the Liu-Hy/GenoTEX HuggingFace mirror; BLADE annotations auto-prepared via blade_bench.
• Sandbox: required for SciCode, ScienceAgentBench, and GenoTEX (likely Docker — several libraries have native dependencies). The union of required libraries across SAB is ~15 scientific packages (deepchem, rdkit, pymatgen, matminer, prolif, neurokit2, scanpy, scikit-learn, …).
• LLM API: BLADE's semantic matcher (GPT-4o-class) and any judge-bearing check need API access.
• Suggested staging: DiscoveryBench + SciCode first (proven in v0.2), then ScienceAgentBench, then GenoTEX (heavy data), then BLADE (evaluator dependency).

Known limitations

• Public-benchmark leakage risk. All five sources are public; frontier models likely saw the source data in training.
• Computational Chemistry over-represented (7 of 30), reflecting ScienceAgentBench's distribution rather than a curation flaw.
• Coverage gaps: no wet-lab/laboratory protocols, no astronomy/astrophysics, no Beakr-domain-specific synthetic tasks, no multi-step / paper-replication tasks.
• Heterogeneous gold fetching — five different mechanisms; GenoTEX data is heavy (82 GB full); BLADE's evaluator depends on an LLM API.
• Designed declaratively before harness integration; check parameters and the newer check types may shift once integration begins.

Out of scope

• Broad scientific Q&A (LAB-Bench, GPQA) — recall, not execution.
• Open-ended hypothesis generation without provided ground truth.
• Multi-paper synthesis.
• Live data sources / network calls during agent execution.
• Long-horizon (multi-day) scientific reasoning.
• Human-in-the-loop interactive science.

License & attribution

Internal-only compilation. For any external distribution (not currently planned), reproduce this table verbatim and include the relevant NOTICE files. ODC-BY and CC-BY are attribution-only; Apache-2.0 requires NOTICE preservation; ODC-By has a separate attribution requirement.

Source	License	Attribution / repo
DiscoveryBench	ODC-BY	Majumder et al., 2024 (arXiv:2407.01725). allenai/discoverybench (original), nhop/discoverybench (reformat).
SciCode	Apache-2.0	Tian et al., 2024 (arXiv:2407.13168). SciCode1/SciCode (problems), scicode-bench/SciCode (eval repo / test_data.h5).
ScienceAgentBench	CC-BY-4.0	Chen et al., 2025 (arXiv:2410.05080). OSU-NLP-Group/ScienceAgentBench.
GenoTEX	CC-BY-4.0	Liu et al., 2025 (arXiv:2406.15341). Liu-Hy/GenoTEX.
BLADE	Code Apache-2.0; Data ODC-By	Gu et al., 2024 (arXiv:2408.09667). github.com/behavioral-data/BLADE.

Changelog

• 0.5.1 — Robustness follow-up scan (AUDIT §9), same policy. DB-002: emptied hypothesis_matches_gold's must_not_contain country list (United Kingdom/Germany/Canada/Australia) — it could reject a correct comparative answer ("primarily the US, more than the UK"); the required United States value plus the judge rubric already gate the correct country, so the brittle list is redundant. No questions, gold hypotheses, or tested properties changed.
• 0.5.0 — Adopted the closed-form-vs-judge scoring policy (documented under Check types). DB-004: removed the bare "significant difference" from hypothesis_matches_gold's must_not_contain, because it is a substring of the correct null result "no significant differences" and the check ANDs the substring layer with the judge (so the judge could not rescue it). The remaining directional tokens still catch confabulated findings; the rubric owns null-result detection. No questions, gold hypotheses, or tested properties changed — only how the check gates.
• 0.4.0 — Conformed to the canonical EvalCase envelope: tasks[] → cases[], each bucketed into input / reference (incl. checks) / fixture / fixture_private / metadata; schema_version → 1.0. Dataset-level sources/curation_criteria/rejected_during_curation retained at top level. No task content changed.
• 0.3.1 — Corrected the SciCode gold-source wording (target values come from test_data.h5 in the eval repo, not the HF dataset); tightened the README.
• 0.3.0 — Added 19 tasks across 3 new sources (ScienceAgentBench, GenoTEX, BLADE); filled the chemistry, biomedical-biology, and statistical-analysis-execution gaps; domain coverage 7 → 15; added 3 check types (program_matches_gold, gtas_pipeline_outputs_match_gold, analysis_decisions_match_gold).
• 0.2.0 — Added 5 tasks (4 DiscoveryBench + 1 SciCode Householder QR) to fill diversity gaps.
• 0.1.0 — Initial 6-task seed (4 DiscoveryBench + 2 SciCode semiconductor physics).