EDB-Bio

EDB-Bio is the evidence-deliberation benchmark from "Do Language Models Know What Compelling Scientific Evidence Looks Like?" (arXiv, code).

Each episode pairs a biological hypothesis with a pool of evidence stubs: real measurements from the same public resource, each shown with its scientific context. The model must report the probability that a held-out measurement supports the hypothesis. The held-out measurement never appears in its own evidence pool, and the amount, relevance and strength of the evidence are controlled during construction. Labels come from the held-out measurement.

Config	Source	Hypothesis	Label is 1 when the held-out measurement shows
tahoe	Tahoe-100M	drug D up/down-regulates gene G in cell line C	DESeq2 padj < 0.05, |log2FC| > 0.5, matching direction
depmap	DepMap 24Q4 + PRISM 24Q2	gene G is a dependency in cell line C	Chronos gene effect ≤ −0.5 and dependency probability ≥ 0.5
gwas	NHGRI-EBI GWAS Catalog	SNP S (risk allele A) influences trait T in a stated direction	p < 5×10⁻⁸ with matching effect direction

Configs

Nine configs in three families. The domain configs hold the episodes; the two testset families are frozen evaluation instances cut from them, so that scores are comparable across runs and papers.

family	config	train	validation	test	what it is
domain	gwas	29,744	256	2,000	the episodes, each with its complete 16-stub context_pool
domain	depmap	29,741	256	2,000	"
domain	tahoe	29,741	256	2,000	"
primary	testset-primary-{gwas,depmap,tahoe}	—	—	2,000	the same 2,000 test episodes, each cut to one sampled k (2–16, seed 0)
sensitivity	testset-sensitivity-{gwas,depmap,tahoe}	—	—	4,352	256 episodes, each repeated 17 times, at k = 0,1,2,…,16

Why the testsets exist: how much evidence a model sees drives its score, so an evaluation that re-samples k is not reproducible. The testsets freeze that choice. testsets/provenance.json records the builder, seed, revision and k ranges.

The pools are physically truncated — no special loader is needed. In both testset families len(context_pool) == k for every row (verified: 2000/2000 and 4352/4352). To run an evidence sweep, load the sensitivity config and group by episode_id; the 17 rows are the 17 values of k.

from datasets import load_dataset
ds = load_dataset("EdisonScientific/EDB-Bio", "testset-sensitivity-gwas", revision="v1.0.0")
# 4,352 rows = 256 episodes x k in 0..16; one row per (episode, k)

Splits

Config	train	validation	test	Evidence pool per episode
gwas	29,744	256	2,000	16 stubs
depmap	29,741	256	2,000	15–16 stubs
tahoe	29,741	256	2,000	up to 32 stubs

Labels are balanced (50% positive). Each episode's context_pool is pre-shuffled; rendering its first k stubs gives nested contexts. The paper trains with k drawn uniformly from 2–16.

Evaluation sets

The paper evaluates on two fixed sets per domain, drawn from the test split, in testsets/. One row is one model call: context_pool is already cut to that row's context length k.

Config	Rows per domain	Episodes	Use
testset-primary-<domain>	2,000	2,000	Headline metrics. Each episode at one k, balanced over 2–16.
testset-sensitivity-<domain>	4,352	256	Evidence-sensitivity analyses. Each episode at every k = 0…16.

Added columns: domain, testset, k (requested length) and n_stubs (actual stub count). The sets are built by scripts/analysis/multidomain/build_static_testset.py at default arguments; a rebuild from this revision is byte-identical. testsets/provenance.json records the arguments and row counts.

Schema

Shared by all three domains:

column	type	description
episode_id	string	stable identifier; the join key across configs
label	int	the answer key: 1 if the held-out measurement supports the claim, else 0
context_pool	list	the evidence stubs revealed to the model, oldest first
target_gene	string	gene the episode is about
target_gene_meta	struct	annotations for that gene
target_direction	string	direction of the claimed effect
negative_type	string	null for positives; for negatives, why the claim fails (see below)
sufficiency_tier	string	rich / moderate / poor — how much supporting evidence the pool was built to contain (see below)

Testset configs add domain, testset, k, n_stubs (k == n_stubs == len(context_pool)).

Domain-specific:

domain	columns
gwas	target_snp, target_snp_meta, target_trait, target_trait_meta, target_allele, target_efo_parent, ood_group, ground_truth
depmap	target_cell, target_cell_id, target_cell_meta, lineage, ood_group, ground_truth
tahoe	target_drug, target_drug_meta, target_cell, target_cell_meta, lineage, lfc_threshold, ground_truth_de

ground_truth / ground_truth_de hold the held-out measurement that label is derived from. They are not shown to a model being evaluated; they are published so the label can be audited.

Columns provided for downstream analysis, not used by the paper's models. negative_type, sufficiency_tier and target_efo_parent are episode strata: the reward environments and the evaluation code do not read them (only the test suite does).

sufficiency_tier

A designed variable, not a measured one (principle P4, Appendix A of the paper). Each episode is assigned a stratum by quota before its pool is built; the pool is then filled with exactly that many supporting relevant stubs and capped there, the rest of the budget going to non-supporting relevant stubs. A poor pool therefore contains zero supporting stubs by construction. Episodes are split in roughly equal thirds. What counts as "supporting" is domain-specific:

domain	what the tier counts
tahoe	conclusive (significant) related stubs: rich ≥ 4, moderate 1–3, poor none. Rich pools average 5.7 conclusive stubs against 0.8 for poor; 88% of all stubs are inconclusive
depmap	evidence supporting a dependency — related cell lines that are themselves dependencies, over-expression, amplification, damaging mutation, drug sensitivity
gwas	genome-wide-significant related stubs, by capping. Because the catalog holds only significant hits, non-significant related evidence is scarce, so poor pools fill largely with unrelated controls

Not to be confused with the paper's "relevance tier". That is a different axis — the expert hierarchy of how directly a stub bears on the target — and it is the one the belief-revision results are reported against.

negative_type

Null for every positive episode. For negatives it records which test the held-out measurement failed, and the categories are domain-specific:

domain	values	meaning
gwas	sub_threshold	held-out association did not reach genome-wide significance
	direction_inconsistent	significant, but the effect allele or direction disagrees with the context
depmap, tahoe	inconclusive	the measurement does not settle the claim
	refuted	the measurement contradicts it

Use

from datasets import load_dataset

ds = load_dataset("EdisonScientific/EDB-Bio", "tahoe", revision="v1.0.0")
test = load_dataset("EdisonScientific/EDB-Bio", "testset-primary-tahoe", revision="v1.0.0")

The reward environments in the code repository render episodes into prompts exactly as in the paper (envs/{gwas,depmap,tahoe}_meta).

Construction

Each domain is built by one seeded script in the code repository (scripts/data_pipeline/build_{gwas,depmap,tahoe_meta}_dataset.py). Every build writes manifest.json recording its parameters and summary statistics. The published files are the canonical version: the GWAS Catalog is updated weekly, so rebuilding gwas from a later download gives a different set of associations. Appendix A of the paper describes the construction, the evidence-relevance hierarchy of each domain, and the design principles.

Known limitation. Splits are assigned i.i.d., so one episode's held-out measurement can appear as an evidence stub in another episode on the other side of the split. This affects 0.6% (Tahoe), 10.1% (DepMap) and 53.2% (GWAS) of test episodes. The paper measures the effect on reported metrics, which is small (at most 0.005 macro Brier for the trained models).

Sources and licences

EDB-Bio is released under CC BY 4.0. It contains values copied or computed from the public resources below. Our licence covers our compilation and derived annotations; upstream content remains subject to its original terms, listed here. All sources were modified (filtered, joined, re-scored, summarised).

Source	Version / files	Upstream licence
NHGRI-EBI GWAS Catalog	All associations (with ontology annotations) and EFO trait mappings, downloaded July 2026	EMBL-EBI Terms of Use
DepMap Public 24Q4 (Broad Institute)	CRISPRGeneEffect, CRISPRGeneDependency, Model, Gene, OmicsCNGene, OmicsExpressionProteinCodingGenesTPMLogp1, OmicsSomaticMutations*, Achilles common-essential / nonessential controls, PortalCompounds	CC BY 4.0
PRISM Repurposing Public 24Q2 (Broad Institute)	Extended Primary Data Matrix, Extended Primary Compound List	CC BY 4.0
Tahoe-100M (Tahoe Therapeutics)	Pseudobulk differential expression, drug and cell-line metadata	CC0 1.0
GTEx v8	Gene median TPM (top-3 tissues per gene, gwas config)	GTEx Portal Data License
Reactome	Ensembl2Reactome.txt (pathway names, gwas config)	CC0

Please also cite the sources:

• Cerezo M, et al. The NHGRI-EBI GWAS Catalog: standards for reusability, sustainability and diversity. Nucleic Acids Res. 2025;53(D1):D998–D1005. doi:10.1093/nar/gkae1070
• DepMap, Broad (2024). DepMap 24Q4 Public. Figshare+. doi:10.25452/figshare.plus.27993248.v1
• DepMap, Broad; Kocak M (2024). Repurposing Public 24Q2. figshare. doi:10.6084/m9.figshare.25917643.v1
• Corsello SM, et al. Discovering the anticancer potential of non-oncology drugs by systematic viability profiling. Nat Cancer 2020. doi:10.1038/s43018-019-0018-6
• Zhang J, et al. Tahoe-100M: A Giga-Scale Single-Cell Perturbation Atlas for Context-Dependent Gene Function and Cellular Modeling. bioRxiv 2025. doi:10.1101/2025.02.20.639398
• GTEx Consortium. The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science 2020;369:1318–1330. doi:10.1126/science.aaz1776
• Ragueneau E, et al. The Reactome Knowledgebase 2026. Nucleic Acids Res. 2025. doi:10.1093/nar/gkaf1223

Attribution required by the sources

None of the upstream providers endorses this dataset.

GTEx. The data used to build this dataset were obtained from the GTEx Portal in July 2026 (GTEx Analysis Release V8, dbGaP accession phs000424.v8.p2). The Genotype-Tissue Expression (GTEx) Project was supported by the Common Fund of the Office of the Director of the National Institutes of Health, and by NCI, NHGRI, NHLBI, NIDA, NIMH, and NINDS. The GTEx-derived fields come from release V8 and do not reflect the most current data available from the GTEx Portal.

GWAS Catalog. Associations are the curated top hits reported by published studies, not full summary statistics. Each row carries its GCST study accession; please also cite the original study where relevant.

Citation

@article{ibis0,
  title   = {Do Language Models Know What Compelling Scientific Evidence Looks Like?},
  author  = {Igoe, Conor and Grand, Gabriel and Koch, Zane and Bou, Albert and Moss, Eli and Hinks, Michaela M. and Rodriques, Samuel G. and White, Andrew and Narayanan, Siddharth},
  journal = {arXiv preprint arXiv:ARXIV_ID},
  year    = {2026}
}