AIAny
Icon for item

forensic-refusal

Provides under-1K JSON agent-trace records documenting model refusal responses and forensic metadata — useful for evaluating refusal-detection, audit pipelines, and safety analysis; small size limits large-scale statistical studies.

Introduction

Why this matters The dataset collects concrete agent interaction traces that capture when and how models refuse requests — a focal point for safety testing, forensic auditing, and tooling that flags problematic refusals. Having curated refusal examples in a consistent JSON agent-trace format shortens the loop for building detectors and audit workflows.

What Sets It Apart
  • Compact, agent-trace centric format: data is serialized as JSON agent traces (structured prompts, model responses, and metadata), so you can plug it directly into trace-based evaluation tools or parsers.
  • Focus on refusal behavior: entries emphasize refusal types and associated context rather than general conversational turns, which makes the set efficient for prototyping refusal classifiers and rule-based auditors.
  • Lightweight for quick iteration: with fewer than 1,000 records, the dataset is small enough to inspect manually and iterate on heuristics or unit tests without heavy compute.
Who It's For & Tradeoffs

Great fit if you need a small, well-structured sample of model refusal cases to prototype detectors, audit pipelines, or demonstration notebooks. Not suited if you require large, diverse corpora for training high-capacity classifiers or for robust statistical evaluation — use this as a seed or validation set rather than a training corpus. Also note limited metadata scope may require augmentation for domain-specific forensic analyses.

Information

  • Websitehuggingface.co
  • OrganizationsHugging Face
  • Published date2026/07/20

Categories

More Items

Hugging Face

Behavior-focused text corpus for LM pretraining, organized into seven Parquet-backed subsets (reasoning, planning, data-science, games, general, format-rewrites, other). Supports streaming, custom sampling, and large-scale dataset pipelines for research and model training.

Hugging Face

Provides multiple Parquet-backed subsets of code problem-solving data (direct answers, chain-of-thought reasoning, and task synthesis) that are streamable and prepared for language-model training and evaluation.

Hugging Face

Provides 30,969 action-conditioned video episodes, each with source MP4, per-frame keyboard control logs, captions, and a COLMAP sparse pose model — intended for research on action-conditioned video prediction, controllable world models, and representation learning.