typesafe/evalsafe-onet

EvalSafe O*NET

150 documents · 7,500 consensus-labeled questions · 9 candidate models. Snapshot: 2026-09-29. Default reference: consensus.

Only questions with an available consensus target and their corresponding documents and final model results are included. The documents are synthetic workplace examples. The reference targets are model-generated, using Astra (gpt-6-astra) and Fable (claude-fable-5-1). The default reference is their consensus.

Load

from datasets import load_dataset

repo = "typesafe/evalsafe-onet"
cases = load_dataset(repo, "cases", split="test")
questions = load_dataset(repo, "questions", split="test")
results = load_dataset(repo, "run_results", split="test")

Use revision="<commit SHA>" to pin the data. If access requires authentication, sign in to Hugging Face first. All three configurations have one test split.

Configuration	Rows	Contents
cases	150	Complete documents, descriptive metadata, and question IDs
questions	7,500	Question/input and independent OpenAI, Anthropic, and consensus labels
run_results	67,500	Final prediction, embedded consensus target, score, cost and usage for each model/question

Each document has 27 noul, 16 score, and 7 choice questions, totaling 4,050 noul, 2,400 score, and 1,050 choice questions. Coverage, scoring definitions, and model summaries are recorded in dataset.json.

Join tables with case_id and question_instance_id; run_id joins model results to the summaries in dataset.json. Fields ending in _json contain JSON-encoded text. Every model-result row includes its reference target.

All 7,500 questions have a consensus target. Both reference sources are available for 7,468 questions; 32 have one available source.

Metrics

• Noul/choice intelligence: 1 - Jensen-Shannon divergence, using base-2 logarithms. This uses divergence, not its square root.
• Score agreement: 1 - abs(model_mean - reference_mean) / rubric_range.
• Overall performance: equal-weight mean of the noul, score, and choice type means. This is a composite of two metrics, not a pure JSD score.

Failed candidates score zero. The final table has 66,528/67,500 valid candidate results. Reference models are excluded from candidate rankings. Jev is reported as jev-1.13.0; the DeepSeek models were served by Together.

Results

Dashed frontiers connect nondominated models and do not predict intermediate configurations. Cost and latency axes are logarithmic; the chart uses all nine candidates.

Model	Performance	Valid results	USD/1,000 questions	sec/question
claude-opus-5	0.94330	7,439/7,500	$19.44560	3.184
gpt-5.6-sol	0.93450	7,438/7,500	$5.67873	2.720
typesafe/jev-1.13.0	0.93386	7,500/7,500	$0.07463	0.141
gpt-5.6-terra	0.91759	7,445/7,500	$2.72939	2.300
claude-sonnet-5	0.90798	7,436/7,500	$7.59240	2.125
claude-haiku-4-5	0.90350	7,468/7,500	$7.91799	12.008
gpt-5.6-luna	0.88725	7,423/7,500	$0.32674	2.281
together/deepseek-ai/DeepSeek-V4-Pro-0813	0.86267	7,245/7,500	$5.60110	9.227
together/deepseek-ai/DeepSeek-V4-Flash-0731	0.83574	7,134/7,500	$0.47691	7.094

These are generated workplace documents and questions, not the original O*NET database tables.

Targets are model-generated rather than human-verified ground truth. This benchmark is not a held-out evaluation of generalization.

License

This dataset is licensed under the Apache License 2.0.