Verify-or-Trust — benchmark data

Data artifacts for the Verify-or-Trust benchmark: does an LLM correctly allocate verification when orchestrating a fallible biology foundation model? The harness (code, Apache-2.0) lives on GitHub; this dataset hosts the inputs it consumes.

At a glance

Field	Value
Primary artifact	substrates/gears_norman.csv
Dataset rows	4,008 decidable (perturbation, gene) edges
Live-verification asset	cells/norman_subset.h5ad with 43,955 cells × 97 readout genes
Code release	Verify-or-Trust v0.1.4
Companion study release	CausalAtlas v0.1.4
License	Apache-2.0

Files

Path	Role	Notes
substrates/gears_norman.csv	Hugging Face datasets-loadable substrate table	One row per edge with predicted effect, held-out label, perturbation regime, and live verification statistics.
cells/norman_subset.h5ad	AnnData asset for the live run_de tool	Sparse matrix backing the Norman 2019 subset used by vot run --real-de cells/norman_subset.h5ad.

Substrate columns

Column	Meaning
perturbation, gene	Edge being evaluated.
fm_log2FC, fm_call	Foundation-model prediction and discretized call.
real_label	Held-out verification label (POSITIVE or TESTED_NEGATIVE).
regime	Perturbation generalization regime.
raw_log2FC, raw_se, raw_q, n_trt, n_cntrl	Measured verification statistics consumed by the public harness.

The shipped CSV may contain blank fm_log2FC values where the source prediction is missing. The v0.1.4 harness emits those as strict JSON null in generated panel JSONL, never as non-standard NaN.

Provenance & license

Derived from public data: Norman et al. 2019 (Science; GEO GSE133344) and GEARS (Roohani et al., Nat. Biotechnology 2024; MIT). These artifacts are our derivation and are redistributable under Apache-2.0.

The benchmark's second substrate, Arc STATE / Tahoe, is not hosted here — Arc's model and outputs are under a non-commercial license, so the harness downloads them from Arc's own repo only when the Tahoe builder is invoked locally (see the GitHub data/README.md).

The GitHub repository includes a machine-readable artifact_manifest.json, JSON schemas, and a public-release validator for strict generated JSONL and substrate checks.

Intended use

• Evaluate whether an agent allocates verification effort when a fallible biology foundation-model prediction is available.
• Build benchmark panels with the public vot panels command and score LLM-free reference baselines with vot baselines.
• Reproduce live differential-expression verification with the included AnnData subset.
• Do not use this dataset as clinical evidence, a diagnostic resource, or a substitute for the original Norman/GEARS datasets.

Use

pip install "git+https://github.com/jang1563/verify-or-trust.git"  # the harness (Apache-2.0)
huggingface-cli download jang1563/verify-or-trust --repo-type dataset --local-dir vot_data
vot panels --substrate-table vot_data/substrates/gears_norman.csv --out panels.jsonl
vot baselines --panels panels.jsonl    # the LLM-free value proof (K1)

Load the substrate as a dataset

from datasets import load_dataset
ds = load_dataset("jang1563/verify-or-trust", "gears_norman")  # the GEARS substrate edges

(The cells/ AnnData is a file for the live run_de tool, fetched via huggingface-cli download.)

Reproducibility checks

The GitHub repository ships schemas and a public release validator:

python3 scripts/validate_public_release.py
vot panels --substrate-table vot_data/substrates/gears_norman.csv --out panels.jsonl
vot baselines --panels panels.jsonl

Expected public-facing invariants:

• substrates/gears_norman.csv loads as the gears_norman Hugging Face dataset config.
• Generated panel JSONL is strict JSON, with missing numeric predictions represented as null.
• cells/norman_subset.h5ad is a sparse AnnData/HDF5 file with 43,955 observations and 97 variables.

Citation

See CITATION.cff, .zenodo.json, and docs/ARCHIVAL_RELEASE.md in the GitHub repository, plus the v0.1.4 GitHub release. A preprint is in preparation.