ajh-oai/genebench-pro-public-package

GeneBench-Pro Public Case Studies

This repository contains public GeneBench-Pro case studies. It is the self-contained package intended for public distribution, including Hugging Face publication.

Package Layout

<repo-root>/
├── .gitattributes
├── README.md
├── LICENSE
├── problems.csv
├── checksums.sha256
├── manifest.json
├── reference_definitions.md
├── reference_grader.py
└── problems/
    └── <eval_id>/
        ├── eval_config.json
        ├── data_files/
        └── report_public.pdf

Each problem directory contains only the release-facing assets:

• eval_config.json: the public task, data-file list, answer schema, ground truth, grader contract, and eval UUID.
• data_files/: the staged files visible to an agent for that problem.
• report_public.pdf: the public scientific case-study report.

Use each problem's eval_config.json for the task, answer schema, reference values, and grading contract.

The ground-truth answers and grader tolerances are intentionally public. This package is for public case studies, reproducibility, and model-analysis examples, not for a hidden-answer leaderboard.

The staged records are released as problem-local benchmark files. The task prompts use the provenance statement: Data provenance is uncertain; use your best scientific judgement.

problems.csv provides a Hugging Face-friendly one-row-per-problem summary table and primary package inventory. manifest.json includes per-problem file paths, SHA-256 checksums, and file sizes; checksums.sha256 provides package-wide integrity information in a standard flat-file format.

Problems

Order	Eval ID	Public title	Domain
0	multiparent_qtl_hmm_lmm	Multi-parent QTL mapping with founder reconstruction	Statistical genetics
1	statgen_cis_mvmr_winnerscurse_scaling_ldaware	LD-aware cis-MVMR with winner's-curse correction	Proteomics / statistical genetics
2	txr1_mtb_causal_sv	Synthetic structural-variant driven tumor therapy decision	Cancer somatic / clinical genomics
3	structural_inversion_subhap_expression_risk	Structural inversion subhaplotype expression and disease risk	Clinical genomics
4	wf_selection	Wright-Fisher selection inference from allele-frequency trajectories	Population genetics
5	hic_sv_masked_loop_strength	Masked Hi-C loop strength under structural variation	3D genomics / structural variation
6	statgen_scrna_ambient_state_eqtl	Ambient RNA and cell-state-aware eQTL mapping	Single-cell statistical genetics
7	carrier_cnv_pseudogene_residual_risk	DRX1 carrier-screening residual risk under CNV and pseudogene calibration	Clinical genomics / carrier screening
8	crispri_casrx_transcript_vs_locus	Transcript-vs-locus CRISPRi decision with CasRx follow-up	Functional genomics
9	popgen_recent_pulse_sexbias	Recent sex-biased admixture pulse inference	Population genetics

The titles in this table are broad showcase titles for the package inventory. The PDF reports use more technical case-study titles on their first pages.

Running a Problem

To run a case study, keep one problem directory intact or copy its eval_config.json together with the files listed in data_files/, preserving those relative paths or updating them in the config. A runner should read eval_config.json, show the agent the task text, make only the staged data_files/ available in the analysis workspace, collect an eval_answer.json, and then grade it with the public grader contract. The config anatomy is: id and eval_uuid identify the problem instance; task describes the run; data_files lists the visible inputs; ground_truth contains the released reference answer; and grader specifies the field-level grading rules and tolerances. In an operational hidden-answer evaluation, only the task text and staged data files would be made available during the run, while ground_truth and grader would stay on the runner side. They are included here because this package is a public case-study and reproducibility release.

Using the Grader Contract

An answer should be a JSON object shaped like:

{
  "answer": {
    "<field_name>": "<value>"
  },
  "reasoning": "<brief method description>"
}

The exact answer fields are problem-specific and are listed in each eval_config.json under ground_truth and grader.config.

To check an answer with the included reference implementation:

python3 reference_grader.py \
  problems/multiparent_qtl_hmm_lmm/eval_config.json \
  path/to/eval_answer.json

The reference grader requires Python 3.10 or newer. The boolean passed field is the authoritative grading decision; score is a diagnostic summary of graded fields and should not be used as a substitute for passed.

See reference_definitions.md for the grader-field definitions and scoring semantics.

License

This public package is available under the MIT License. See LICENSE.