punyajain/epistemic-debugging-benchmark

Epistemic Debugging Benchmark (v2)

A benchmark for diagnosing failed experiments and engineering faults. Each item gives a messy evidence summary plus competing hypotheses, and the model has to work out which cause the evidence actually supports, which near-miss explanations it rules out, and which follow-up check would tell them apart. It tests reasoning about evidence; it is not a fault-code lookup.

Project: Epistemic-Debugging-Benchmark on GitHub. Baseline model: punyajain/epidebug-mm-v3.

Splits

split	rows
train	1177
validation	301
test	337

The test split holds out whole failure families (BIO-SAMPLE-HOLD, CHEM-HPLC-DEGAS, ELEC-PCB-ELECTROMIGRATION, MAT-CONCRETE-CHLORIDE, MECH-BOLT-PRELOAD, MFG-TOOLWEAR, ROB-PREMATURE-OPEN, SEN-LOCKIN-PHASE, SW-TEMPORAL-LEAK, THM-FOULING), so no paraphrase or variant of a test family appears in train. Scores on test measure generalization to new failure mechanisms.

Domains

domain	rows
mechanical	174
electrical	164
robotics	166
chemistry	74
materials	108
software_ml	144
fluids_thermal	110
manufacturing_industrial	553
biology_lab	155
sensors	143
physics_priors	24

Sources

source_type	rows
synthetic_template	1275
expert_authored_synthetic	62
synthetic_schema_harvest	89
harvested_public_ai4i	296
synthetic_physics_prior	24
expert_authored_synthetic_paraphrase	65
real_pipeline_sim_injected_fault	4

Most rows are synthetic, generated from expert-written failure families with physics priors and hard near-miss distractors. harvested_public_ai4i rows come from the public AI4I 2020 predictive-maintenance dataset (UCI, CC BY 4.0). real_pipeline_sim_injected_fault rows are robot pick-and-place runs from a simulated UR5 pipeline with injected gripper faults. Every row has a source_type field so you can filter.

Fields

id, group_id (failure family), source_type, domain, subdomain, failure_category (one of 4), difficulty (easy/medium/hard/expert), title, evidence, physics_priors, competing_hypotheses, plus gold labels for the cause, the ruled-out hypotheses, and the discriminating follow-up.

Scoring

Report accuracy and macro-F1 on failure_category, top-1 accuracy at ranking the gold hypothesis (shuffle the options so the gold one isn't always first), and follow-up selection accuracy, all on the test split.

Baseline (epidebug-mm-v3, CPU, frozen MiniLM + small heads)

metric	validation	test
category accuracy	0.548	0.496
category macro-F1	0.513	0.501
hypothesis ranking (shuffled)	0.578	0.632
follow-up selection	0.492	0.362

Much lower than v1 of this benchmark, on purpose: the near-miss distractors are harder and test families are held out. Lots of headroom.

Limitations

The data is mostly synthetic and template-based, so there will be repeated patterns. Labels are author-assigned, not peer-reviewed. The evidence is text summaries; raw sensor, image and CAD data are not included. Don't use the benchmark for safety certification.

Citation

@misc{jain2026epidebug,
  title={Epistemic Debugging Benchmark},
  author={Punya Malaiya Jain},
  year={2026},
  url={https://github.com/Punyajain-cmd/Epistemic-Debugging-Benchmark}
}