OpenDecisionMix-1M — Jev Predictions

OpenDecisionMix-1M contains 1,000,000 successful structured predictions from TypeSafe's jev-1.13.0 API. Each row combines an upstream decision episode, its Jev-compatible request, the original comparison label, and the complete Jev response returned for that request.

This release is a general decision-modeling corpus. It is not a clinical dataset and contains no MIMIC data.

Release summary

Item	Value
Rows / unique row_id values	1,000,000
Parquet shards	51
Train	957,323 rows / 48 shards
Validation	18,896 rows / 1 shard
Test	23,781 rows / 2 shards
Jev model	jev-1.13.0
Jev input tokens	2,037,480,148
Jev output tokens	59,008,331
Estimated Jev input cost	$85.574166216
Rows with a Jev error	0

The cost is the per-row recorded estimate at $0.042 per million input tokens. It is not an invoice or a claim about account credits remaining.

What one row represents

One row is one real Jev request/response pair. Important columns include:

• state_json, question_text, criteria_json, primitive, and request_json: the exact structured decision request;
• gold_label, gold_score, gold_json, and label_source: the original human, automated, or programmatic comparison target;
• jev_prediction, jev_answer_type, jev_confidence, and jev_probabilities_json: the normalized Jev prediction;
• jev_answer_json and jev_raw_response_json: the complete answer and raw API response retained as JSON strings;
• jev_input_tokens, jev_output_tokens, jev_estimated_cost_usd, jev_latency_ms, and jev_completed_at: request-level execution metadata;
• task, source, pinned revision, license, transformation, and split provenance.

Jev outputs are model predictions, not ground-truth labels. The original labels are retained so users can measure agreement, filter or calibrate the predictions, train behavior-cloning or decision heads, build reward-model and RL experiments, or evaluate another structured decision model.

Decision types

Jev primitive	Rows
choice	579,457
noul	281,112
score	139,431

Source mix

Source	Derived rows
nvidia/Nemotron-SFT-Agentic-v2	357,734
nvidia/Aegis-AI-Content-Safety-Dataset-2.0	150,000
wandb/RAGTruth-processed	147,790
nvidia/HelpSteer3	116,826
nvidia/Open-SWE-Traces	100,000
nvidia/HelpSteer2	88,271
nvidia/When2Call	26,296
PolyAI/banking77	13,083

Task mix

Task family	Rows
customer_service_next_action	88,427
multi_attribute_response_score	88,271
claim_span_grounding	81,528
agent_next_action_type	79,238
agent_workload_routing	74,030
response_helpfulness_score	51,160
specific_tool_usage	44,444
customer_service_tool_selection	41,573
tool_argument_completeness	37,233
tool_selection	37,233
conversation_domain_routing	25,097
pairwise_response_preference	25,225
hallucination_type_presence	24,236
prompt_safety	23,644
rag_task_routing	17,790
tool_or_text_action	14,349
customer_intent_routing	13,083
principle_fulfilment	12,194
rag_response_quality	12,118
response_grounding	12,118
safety_category_presence	117,554
patch_presence	11,111
trace_completion_shape	11,111
trace_disposition	11,111
trace_outcome	11,112
trace_tool_usage	11,111
response_safety	8,802
tool_response_preference	8,295
when_to_call_tool	3,652
edit_quality_preference	3,150

Validation and splitting

The final audit verified:

• exactly 1,000,000 rows and 1,000,000 unique row_id values;
• all 51 expected Parquet shards and the declared split totals;
• byte-equivalent logical content for every immutable seed and Jev field when compared row by row with the checkpointed rollout;
• no jev_error values and nonempty model, type, prediction, answer, raw response, and completion fields;
• raw-response agreement with normalized answers, models, and token usage;
• valid per-row and aggregate usage/cost arithmetic;
• no group_id or upstream source-row identity crossing output splits.

The initial physical partition contained connected source/group components in more than one split. The release moved 12,654 rows between splits without changing any request, prediction, label, or provenance field. Exact split sizes were preserved. This prevents the same source row or decision group from appearing in more than one split.

Models trained on this set

Both use the source labels only. No Jev answers were used in training.

• Qwen3.5-4B-Router-1M: an agent's next step and tool choice, 86.5% vs Jev 1.13 80.6% on 480 held-out rows.
• Qwen3.5-4B-Judge-1M: RAG grounding and safety checks, 89.5% vs Jev 1.14 82.2% on 4,035 rows of an earlier test file. Jev still ranks real PubMed answers above fake ones better (916 vs 862 of 1,000).

Limitations

• The predictions have not been independently established as correct or calibrated. Treat Jev confidence values as model outputs.
• All jev_request_id values are empty because the API responses did not expose the expected request identifier. This does not indicate a missing prediction; each row retains its full successful response and completion metadata.
• Upstream corpora can contain duplicated content, noisy or conflicting labels, third-party model outputs, and potentially disturbing safety examples.
• This mixture is broad and heavily weighted toward agent, safety, RAG, and customer-service decisions. It should not be presented as clinical evidence.
• JSON payloads are stored as strings for a stable cross-task Parquet schema.

Sources and licensing

Every row records source_dataset, source_revision, source_license, and source_url. The aggregate uses license: other because provenance and terms are mixed. Most source rows declare CC BY 4.0; RAGTruth-derived rows are marked mit-with-upstream-review-required and require review of upstream material and model-output terms. Users are responsible for following each recorded source's license and any model/API terms that apply to their use.

Repository layout

The release is deliberately flat: a root README.md plus train-*.parquet, validation-*.parquet, and test-*.parquet files. It includes no compiler, generation code, audit scripts, logs, secrets, failed requests, or checkpoint directories.