Datasets:
bench: save per-decision probabilities and add them to the eval-results PR
score_with_jobs.py already scores every decision but keeps only the summary. This PR:
- writes
predictions.jsonlnext toresults.json: one line per decision with case id, workflow, question id, type and the normalized probabilities over the gold options (the vectorscore_onescores, after the yes/no to true/false mapping); - has
submit_results.pyadd that file to the model-repo PR aseval/typed-decisions/predictions.jsonlwhen it is present next to the results file (--predictionspoints elsewhere).
Scoring is unchanged. Replaying a published Jev run (OpenDecider's whole-case file, which their benchmarks README lists as Jev 1.13 through TypeSafe's API) through both versions with vLLM, HTTP and datasets mocked gives a byte-identical results.json (accuracy 0.7305), and re-scoring predictions.jsonl gives the same accuracy, KL, Brier and ECE.
The tie rule (first-listed option vs split credit) is left for a separate change; with the probabilities saved it can be re-applied to any run.
Context: https://huggingface.co/posts/tardellirs/145368791310880
A question that came up in the tie discussion on the explorer's post (not a change to this PR): does the generator keep the three per-option samples behind each gold distribution, or only the means?
With samples on a 0.01 grid, the 35 exact two-way ties in test look like they are broken by float summation order before rounding, so label cannot be checked against the parquet alone. If the samples or the unrounded means exist, publishing them would settle which option label should take on those decisions, and which tie rule the scorer should use.