Aadharsh/cats-bench-kev

CaTSBench Kev

CaTSBench Kev converts numeric time series and human captions from mhfisher/CaTSBench into labelled TypeSafe System One requests. It is an evaluation dataset for Kev, Jev, and other decision models. It contains both questions about the visible series and questions about a hidden suffix.

The validation split has 459 records, 950 questions, and 402 source-series groups. Each record contains one shared state and one or more Choice, Noul, or Score questions. There are no training or test records in this release.

Construction

The source dataset is pinned to revision 63e35cded0d5a8e3a3fd87d487da267cf873a846. The builder sampled 500 test series with human captions using seed catsbench-pilot-v1.

Qwen/Qwen3.8-27B at revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 selected applicable question templates. The model did not produce labels. Deterministic Python functions calculated every label from the numeric series. Forecast records expose only the prefix used by the question.

The builder removed duplicate states and excluded two underpowered templates, future_change and observed_monotonic_decrease. It then applied Kev's training-context admission rule with the released Kev-4B tokenizer.

Files

• development.jsonl contains the benchmark records.
• manifest.json pins sources, revisions, hashes, context limits, and the construction protocol.
• proposals.jsonl preserves every raw Qwen template-selection response.
• audit.json contains source-backed label, prefix, leakage, duplication, and balance checks.
• The other JSONL partitions are empty so the folder can also be loaded as a Kev frozen suite.

Record format

{
  "state": "{\"domain\":\"agriculture\",\"times\":[0,1,2],\"values\":[96.46,102.01,104.59]}",
  "questions": {
    "direction": {
      "type": "choice",
      "instructions": "From the first observation to the last, did the value decrease, stay stable, or increase?",
      "criteria": {
        "decrease": "decrease",
        "stable": "stable",
        "increase": "increase"
      },
      "label": "increase",
      "src": "catsbench_observation"
    }
  }
}

group_id keeps observation and forecast records derived from the same source series together. Use it when splitting or bootstrapping the dataset.

Validation

The audit recomputes every label from the pinned source archive and verifies that each forecast state is a strict prefix. It reports no invalid records, duplicate states, caption leakage, or metadata-statistic leakage. The forecast-direction family has only one stable example, so per-class results for that family are preliminary.

Initial results

The initial run uses jaredpalmer/kev-4b at revision 1da696f7938f77c4cdf5471e92fd342baff41778. It evaluates every record without truncation or rejection.

Split	Questions	Accuracy	NLL	ECE
All	950	62.8%	0.855	0.093
Observation	859	66.2%	0.771	0.060
Forecast	91	30.8%	1.655	0.434

Results by question family show where the model succeeds and fails. The majority column is the accuracy obtained by always predicting that family's most common label.

Question family	Questions	Kev-4B	Majority
Forecast direction	38	26.3%	57.9%
Forecast new high	23	43.5%	52.2%
Forecast new low	30	26.7%	63.3%
Observation direction	208	85.6%	52.4%
Observation maximum position	317	48.3%	41.3%
Observation relative change	303	70.6%	35.6%
Observation second-half comparison	31	77.4%	51.6%

Kev-4B performs above the majority baseline on every observation family and below it on every forecast family. The forecast result also has poor calibration, with an ECE of 0.434. This first run therefore supports using observation and forecast questions as separate benchmark sections.

The complete outputs are in baselines/kev-4b/report.json and baselines/kev-4b/summary.json. Reproduce the run from a Kev checkout with:

uv run python -m kev.benchmark \
  --run jaredpalmer/kev-4b \
  --suite evals/external/catsbench-v1 \
  --out runs/kev-4b-catsbench-v1 \
  --device cuda

License

CaTSBench and this derived dataset use the MIT license.