Datasets:

mkzero
/

agent-decision-30k

language:
  - en
  - zh
license: apache-2.0
task_categories:
  - text-classification
  - feature-extraction
  - zero-shot-classification
tags:
  - agent
  - decision-making
  - cross-encoder
  - reranker
  - modernbert
  - tool-use
size_categories:
  - 10K<n<100K

Agent-Decision-30k: Multi-Task Operational Decisions Dataset

Agent-Decision-30k is a curated dataset of 30,456 high-fidelity operational decisions for training and benchmarking System-1 (fast-thinking) decision engines and cross-encoders.

Dataset Structure

Each sample is formatted as a JSON object:

{
  "context": "Detailed observation, system logs, environment state, or DOM context",
  "instruction": "Target goal, user intent, or operational constraint",
  "candidates": {
    "action_a": "First candidate option",
    "action_b": "Second candidate option",
    "action_c": "Third candidate option"
  },
  "gold": "action_a",
  "domain": "browser_dom | api_dispatch | ecom_brand | ecom_specs | ecom_category | legal_clause"
}

Domains Covered

Browser DOM Automation: XPath selection, button clicking, form field focus.
API Tool Routing: Function calling and handler dispatch under strict argument constraints.
E-Commerce Brand Extraction: Canonical brand attribution across messy product titles.
E-Commerce Specs & Attributes: Multi-attribute constraint satisfaction.
E-Commerce Category Prediction: Hierarchical taxonomic placement.
Legal & Contract Alignment: Clause matching and regulatory compliance verification.

Intended Use

Designed to train sub-15ms cross-encoders (such as FastDecider-149M) for reliable agent dispatch.