sparrow8i8's picture
add output-overlap verification (13-gram scan: no seed text in released corpora)
3065467 verified
|
Raw
History Blame Contribute Delete
2.31 kB
metadata
dataset_info:
  features:
    - name: seed_text
      dtype: string
    - name: tier
      dtype: string
    - name: animal_relevant
      dtype: bool
    - name: category
      dtype: string
  splits:
    - name: train
      num_bytes: 1379238368
      num_examples: 893935
  download_size: 572016561
  dataset_size: 1379238368
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
extra_gated_heading: 'Access request: research seed corpus'
extra_gated_prompt: >-
  This dataset contains de-identified excerpts of public online discussions
  about animal welfare, provided solely as topic-grounding seed material for
  research. Structured identifiers have been removed, but the free text
  originates from third-party authors and is NOT licensed for redistribution. By
  requesting access you agree to the conditions below.
extra_gated_fields:
  I will use this dataset for research or analysis only: checkbox
  I will not redistribute or republish the dataset or substantial excerpts of it: checkbox
  I will not attempt to re-identify the authors of any text: checkbox
  I understand the text originates from third-party authors and my use must comply with applicable law and platform terms: checkbox
  Intended use (one line): text

Animal-compassion seed corpus (gated)

893,935 de-identified excerpts of public online discussion about animal welfare (columns: seed_text, category, tier, animal_relevant). Used by CaML's synthetic data pipelines as TOPIC grounding only: generation prompts inject short snippets as inspiration and instruct the model never to copy text or adopt stance. Generated corpora contain no text from this dataset.

Access is gated: use for research/analysis only, no redistribution, no re-identification attempts.

Provenance and output verification (updated 16 July 2026)

This corpus is used strictly as topic-grounding seed material: generation prompts inject short snippets as inspiration and instruct the model never to copy text or adopt stance. This is verified, not asserted: a 13-word n-gram scan across all 893,935 seeds against CaML's released synthetic corpora found overlap in only 10 seeds (0.001%), every one an official treaty name (e.g. the full name of CITES) or standard statutory phrasing, and none of them user discussion text.