README.md

metadata

pretty_name: Intent and Domain Routing — Jev 130K
license: other
language:
  - en
task_categories:
  - text-classification
  - text-ranking
  - reinforcement-learning
tags:
  - agents
  - tool-calling
  - structured-prediction
  - preference-data
  - reward-modeling
  - jev
  - typesafe
configs:
  - config_name: default
    data_files:
      - split: train
        path: train-*.parquet
      - split: validation
        path: validation-*.parquet
      - split: test
        path: test-*.parquet

Intent and Domain Routing — Jev 130K

130,000 validated typed decisions for modern AI systems. Intent, conversation-domain, RAG-task, and agent-workload routing 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

• intent and domain classifiers
• model/workload routers
• RAG task routing
• multilingual routing failure analysis

Release summary

Item	Value
Rows / unique row IDs	130,000
Unique decision groups	129,976
Unique upstream source rows	130,000
Train	119,467
Validation	2,829
Test	7,704
Jev model	jev-1.13.0
Jev input tokens	242,740,155
Jev output tokens	16,016,902
Recorded estimated input cost	$10.195086510
Rows with a Jev error	0

Quick start

from datasets import load_dataset

dataset = load_dataset("MaziyarPanahi/Jev-Intent-Domain-Routing-130K")
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
agent_workload_routing	74,030
conversation_domain_routing	25,097
rag_task_routing	17,790
customer_intent_routing	13,083

Decision primitives

Primitive	Rows
choice	130,000

Source mix

Source	Rows
nvidia/Nemotron-SFT-Agentic-v2	74,030
nvidia/HelpSteer3	25,097
wandb/RAGTruth-processed	17,790
PolyAI/banking77	13,083

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 130,000 rows with directly comparable source labels, Jev/source-label agreement is 92.52%. 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
agent_workload_routing	74,030	100.00%	Descriptive Jev/source-label agreement only
conversation_domain_routing	25,097	72.65%	Descriptive Jev/source-label agreement only
customer_intent_routing	13,083	78.56%	Descriptive Jev/source-label agreement only
rag_task_routing	17,790	99.67%	Descriptive Jev/source-label agreement only

High-confidence (jev_confidence >= 0.90) disagreements: 3,498.

Validation and splitting

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

• exactly 130,000 rows and unique row IDs;
• one priority only: intent_domain_model_router;
• 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

• The source/domain taxonomy is heterogeneous across upstream datasets.
• The HelpSteer3 preview showed weak multilingual recall for Jev routing.
• 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.