MaziyarPanahi/Jev-Policy-Verification-180K

Policy Verification — Jev 180K

180,000 validated typed decisions for modern AI systems. Response scoring, preference, helpfulness, and principle-fulfilment decisions.

General-purpose dataset: this release contains no clinical or medical data. Jev outputs are model predictions, not ground truth.

Each row pairs an upstream reference label with a pinned jev-1.13.0 prediction, confidence, structured answer, and provenance. The flat Parquet layout works directly with Hugging Face datasets, DuckDB, Polars, or pandas.

What it is good for

• reward-model and preference-model research
• policy/principle adherence heads
• response-quality ranking and calibration
• teacher-human disagreement studies

Release summary

Item	Value
Rows / unique row IDs	180,000
Unique decision groups	45,281
Unique upstream source rows	84,007
Train	171,719
Validation	5,690
Test	2,591
Jev model	jev-1.13.0
Jev input tokens	221,426,328
Jev output tokens	4,386,328
Recorded estimated input cost	$9.299905776
Rows with a Jev error	0

Quick start

from datasets import load_dataset

dataset = load_dataset("MaziyarPanahi/Jev-Policy-Verification-180K")
row = dataset["train"][0]

print(row["state_preview"])
print(row["question_text"])
print(row["jev_prediction"], row["jev_confidence"])

For a private repository, authenticate first with hf auth login. Once the repository is public, the same snippet works without authentication.

How to read a row

Need	Columns
Stable identity and split	row_id, group_id, output_split
Decision task	priority, task_family, primitive
Model input	state_json, state_preview, question_text, criteria_json
Upstream reference	gold_label, gold_score, gold_json, label_source
Jev result	jev_prediction, jev_confidence, jev_probabilities_json, jev_answer_json
Full receipt	request_json, jev_raw_response_json, token usage, latency, request ID
Provenance	source dataset, pinned revision, row ID, license, URL, transformation

Decision primitives are intentionally simple:

• choice: one categorical label plus a probability distribution;
• noul: a bounded value from 0 to 1;
• score: a structured numeric assessment.

Task mix

Task family	Rows
multi_attribute_response_score	88,271
response_helpfulness_score	51,160
pairwise_response_preference	25,225
principle_fulfilment	12,194
edit_quality_preference	3,150

Decision primitives

Primitive	Rows
score	139,431
choice	28,375
noul	12,194

Source mix

Source	Rows
nvidia/HelpSteer3	91,729
nvidia/HelpSteer2	88,271

Every row retains the pinned source revision, source license, source row ID, transformation, group ID, original label, complete Jev request, complete raw response, normalized prediction, probability distribution or score, usage, and latency.

Descriptive agreement audit

Across the 40,569 rows with directly comparable source labels, Jev/source-label agreement is 74.60%. This is a behavior audit, not an accuracy or benchmark claim; upstream labels can be noisy or underspecified.

Task family	Evaluated	Agreement	High-level interpretation
edit_quality_preference	3,150	91.05%	Descriptive Jev/source-label agreement only
pairwise_response_preference	25,225	71.77%	Descriptive Jev/source-label agreement only
principle_fulfilment	12,194	76.21%	Descriptive Jev/source-label agreement only

High-confidence (jev_confidence >= 0.90) disagreements: 571.

Validation and splitting

The release uses repaired component-safe splits. Its independent audit verifies:

• exactly 180,000 rows and unique row IDs;
• one priority only: policy_verification;
• no decision group or upstream source row crosses splits;
• nonempty model, prediction, answer, raw response, completion, and provenance fields;
• zero jev_error values;
• valid choice probability distributions and in-schema predictions;
• matching normalized and raw Jev answers;
• pinned source revisions and flat, viewer-friendly Parquet shards.

Important limitations

• Absolute Jev scores should be treated as soft supervision.
• Preference ties and fine-grained scores are intrinsically noisy.
• Confidence values are uncalibrated model outputs.
• Exact-request duplicates are retained where they reflect repeated or conflicting upstream evidence; consumers may aggregate or deduplicate by request_json, group_id, or source identity.
• The card declares license: other: every row records source-specific license and provenance, and users are responsible for reviewing the applicable upstream and provider terms.
• This is a general decision-modeling dataset, not a clinical dataset.