MaziyarPanahi/Jev-Agent-Trace-Audit-100K

Agent Trace Audit — Jev 100K

100,000 validated typed decisions for modern AI systems. Software-agent trace outcome, completion, tool-use, patch, and disposition 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

• agent trajectory scoring and trace auditing
• tool-use and patch-presence classifiers
• software-agent outcome modeling
• failure analysis and reward shaping

Release summary

Item	Value
Rows / unique row IDs	100,000
Unique decision groups	5,556
Unique upstream source rows	5,556
Train	97,156
Validation	1,638
Test	1,206
Jev model	jev-1.13.0
Jev input tokens	309,660,390
Jev output tokens	3,033,782
Recorded estimated input cost	$13.005736380
Rows with a Jev error	0

Quick start

from datasets import load_dataset

dataset = load_dataset("MaziyarPanahi/Jev-Agent-Trace-Audit-100K")
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
specific_tool_usage	44,444
trace_outcome	11,112
patch_presence	11,111
trace_completion_shape	11,111
trace_disposition	11,111
trace_tool_usage	11,111

Decision primitives

Primitive	Rows
noul	77,777
choice	22,223

Source mix

Source	Rows
nvidia/Open-SWE-Traces	100,000

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 100,000 rows with directly comparable source labels, Jev/source-label agreement is 67.50%. 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
patch_presence	11,111	49.39%	Descriptive Jev/source-label agreement only
specific_tool_usage	44,444	73.74%	Descriptive Jev/source-label agreement only
trace_completion_shape	11,111	100.00%	Descriptive Jev/source-label agreement only
trace_disposition	11,111	36.53%	Descriptive Jev/source-label agreement only
trace_outcome	11,112	26.62%	Descriptive Jev/source-label agreement only
trace_tool_usage	11,111	100.00%	Descriptive Jev/source-label agreement only

High-confidence (jev_confidence >= 0.90) disagreements: 813.

Validation and splitting

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

• exactly 100,000 rows and unique row IDs;
• one priority only: agent_trace_auditor;
• 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

• Resolver outcomes are external evaluation signals, not semantic correctness proofs.
• Repository and task distributions are specific to the upstream trace corpus.
• 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.