Hugging Face's logo

Datasets:
zoom-ai
/
contested-fast-decisions

Tasks:
Text Classification
Modalities:
Text
Formats:
json
Languages:
English
Size:
< 1K
ArXiv:
Tags:
classification
multi-task
decision-model
routing
adversarial-filtering
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
contested-fast-decisions / README.md
sileixu's picture
sileixu
initial commit
43c12ac unverified
preview code
|
Raw History Blame Contribute Delete
12.7 kB
metadata 
license: apache-2.0
language:
  - en
task_categories:
  - text-classification
pretty_name: Contested Fast Decisions
size_categories:
  - n<1K
tags:
  - classification
  - multi-task
  - decision-model
  - routing
  - adversarial-filtering
source_datasets:
  - fastino/fast-decisions
configs:
  - config_name: default
    data_files: data/train.jsonl
    default: true

Contested Fast Decisions: A Benchmark of Contested Operational Decisions

Contested Fast Decisions is a benchmark of 818 operational decisions (routing, triage and classification) over 696 texts in 17 domains, on which current models do not reliably agree. Every label was re-checked. It is derived from the development split of fastino/fast-decisions.

Leaderboard

Decision models, queried through their API with one choice question per single-label decision and one yes/no noul question per label for multi-label decisions, given only the task name and candidate labels (no label descriptions).

Model Accuracy Single-label Multi-label
Zoom Decisions Pro 73.7% 73.7% 73.8%
Zoom Decisions 69.3% 67.9% 74.3%
LiquidAI d1-ultra 66.5% 67.6% 62.8%
Perplexity Decision (v1.2 27B) 66.3% 64.4% 72.7%
JEV (jev-1.14.0) 66.0% 68.5% 57.4%
StartLux Decision 27B Instant 65.2% 66.6% 60.1%
Perplexity Decision (v1.1 27B) 63.2% 64.4% 59.0%
OpenAI Decisions API (gpt-6-luna) 62.7% 63.5% 60.1%
JEV (jev-1.13.0) 62.6% 65.5% 52.5%
Microsoft Decision 1 61.0% 64.4% 49.2%
vLLM-SR d3 (26B) 60.6% 61.4% 57.9%
LiquidAI d1 60.1% 60.3% 59.6%
StartLux Decision 9B Instant 60.1% 63.1% 49.7%
Cloudflare Clef 57.2% 58.4% 53.0%
StartLux Decision 4B Instant 56.2% 58.3% 49.2%
vLLM-SR d3-flash (8B) 56.0% 57.2% 51.9%
vLLM-SR d3-mini (5B) 55.7% 59.1% 44.3%
Bespoke Nimble 9B v3 54.4% 60.9% 31.7%
StartLux Decision 2B Instant 49.9% 54.0% 35.5%
Cloudflare Clef-Flash 46.2% 51.3% 28.4%
vLLM-SR d3-nano (2B) 45.8% 49.1% 34.4%
GLiNER2.5-Decide 45.1% 55.4% 9.3%

How it was built

Starting from the fastino/fast-decisions dev split (1,700 texts, 2,900 decisions):

  1. Keep the contested decisions. Eight fixed filter models answer every decision (gpt-6-luna, Gemma 4 31B-it, gpt-5.6-luna, JEV (jev-1.13.0), Qwen3.8-27B, Qwen3.6-27B, Qwen3.5-9B, Qwen3.6-35B-A3B; all without extended thinking). Decisions on which all eight give the same answer are discarded; the 983 on which they do not all agree are kept.
  2. Relabel them. The original labels are unreliable (for example, a tennis recap with a clear winner labelled draw), so each kept decision is labelled again, independently, by gpt-6-sol (default reasoning) and Claude Opus 5 (default thinking), from the same prompt and without seeing the original label or any model's answer. They agree on 686; that label is used.
  3. Resolve disagreements. The remaining 297 (the labellers disagreed, or Opus 5 declined to answer) were reviewed one at a time by Claude Opus 5.5, which set the label for 132 and judged 165 too ambiguous to label; those are dropped.

Because every decision was contested, each is answered wrong by at least one filter model. The filter models only select the data; any model can be evaluated on exactly these decisions.

424 of the 818 labels (51.8%) differ from the original dataset. Each decision keeps its original label in original_label.

Data format

A single file, data/train.jsonl, with one JSON object per text. Only the text's hard decisions are included, so a text can carry fewer decisions than in the original dataset.

{
  "id": "sports_recap/2",
  "category": "sports_recap",
  "input": "<free text>",
  "output": {
    "classifications": [
      {
        "task": "sport",
        "true_label": [
          "basketball"
        ],
        "labels": [
          "football",
          "basketball",
          "baseball",
          "soccer",
          "tennis",
          "cricket",
          "motorsport",
          "other"
        ],
        "multi_label": false,
        "original_label": [
          "other"
        ],
        "label_source": "labeller_agreement"
      }
    ]
  }
}
Field Meaning
id <category>/<line>: the text's line number (0-based) in fastino/fast-decisions <category>.jsonl; the category is also in the category field
input the text, unchanged from the source dataset
task, labels, multi_label the decision and its full candidate label set, unchanged from the source dataset
true_label the verified gold label(s); always a list
original_label the label in fastino/fast-decisions
label_source labeller_agreement (gpt-6-sol and Claude Opus 5 agreed; 686 decisions) or adjudication (set by Claude Opus 5.5; 132 decisions)
Domain Texts Decisions
ticket_route 77 121
email_triage 62 101
restaurant_review 79 86
product_feedback 71 78
screen_tags 53 64
clinic_request 41 47
sports_recap 37 44
benefits_request 43 44
news_topic 35 35
support_topic 34 34
travel_request 34 34
support_intent 30 30
banking_intent 28 28
paper_field 23 23
agent_handoff 18 18
review_sentiment 16 16
document_type 15 15
Total 696 818

635 decisions are single-label and 183 are multi-label (product_area, aspects, genres).

Loading 

from datasets import load_dataset

ds = load_dataset("zoom-ai/contested-fast-decisions", split="train")                # all domains
email = ds.filter(lambda r: r["category"] == "email_triage")

Evaluation

Exact match per decision: a prediction is correct only if the predicted label set equals true_label (compared as sets). Report plain accuracy over all 818 decisions; per-domain counts are small (15 to 121), so per-domain averages are noisy.

License and attribution

Apache 2.0, following the source dataset. If you use this dataset, please cite it:

@misc{xu2026contestedfastdecisions,
      title={Contested Fast Decisions: A Benchmark of Contested Operational Decisions},
      author={Silei Xu, Linyong Nan},
      year={2026},
      publisher={Hugging Face},
      howpublished={\url{https://huggingface.co/datasets/zoom-ai/contested-fast-decisions}},
}

Texts, tasks and candidate labels come from fastino/fast-decisions; please cite it as well:

@misc{zaratiana2025gliner2efficientmultitaskinformation,
      title={GLiNER2: An Efficient Multi-Task Information Extraction System with Schema-Driven Interface},
      author={Urchade Zaratiana and Gil Pasternak and Oliver Boyd and George Hurn-Maloney and Ash Lewis},
      year={2025},
      eprint={2507.18546},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2507.18546},
}

Last updated: 10/16/2026