Maryamp/welfare-facts

WelfareFacts

35 verifiable numeric facts about farm animals, built to test whether LLM errors lean towards making current practice look kinder.

• 28 EU welfare-law values: Directives 1999/74, 2008/120, 2007/43 and 2008/119, and Regulations 1/2005 and 1099/2009.
• 7 FAOSTAT 2024 world slaughter counts.

Files

File	Rows	Contents
facts.jsonl	35	id, question, answer, unit, flatter (error direction that flatters the status quo: higher/lower), source, quote (verbatim passage from the source), tol (0 = exact; 0.2 = within 20%), category
model_answers.csv	175	answers from 5 models via NVIDIA NIM (gemma-4-31b-it, gpt-oss-20b, muse-glimmer-30b, nemotron-3-super-120b, nemotron-3.5-lightning-30b): answer, confidence, correct, flattering
count_errors.csv	17	wrong slaughter counts with the latest FAOSTAT year each would fit (a check for out-of-date training data)
faostat_world_slaughtered.csv	448	FAOSTAT world "Producing Animals/Slaughtered", 1961–2024, for the 7 species (FAO, CC BY 4.0)

Results (pilot)

• Every model understates slaughter. 16 of 17 wrong counts are too low (p = 0.0003); GPT-OSS-20B gives 1.6 billion chickens vs 78.5 billion.
• Legal errors have no consistent direction: 34 of 60 flatter (p = 0.37).
• Accuracy ranges from 17% to 83%.

Sources and licences

• EU texts: © European Union, reused under Commission Decision 2011/833/EU.
• FAOSTAT: FAO, CC BY 4.0.
• Model answers: subject to each model provider's terms.
• This dataset: CC BY 4.0.

Code, full report and figures: https://github.com/Mariihmp/ai-animal-welfare-evals/tree/main/welfare-facts

Datasets, prompts and analysis code were developed with LLM assistance. All results are pilot-scale and have not yet been reviewed by domain experts.

Citation

@misc{welfare_facts_2026,
  author = {marhmpa},
  title  = {WelfareFacts},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Maryamp/welfare-facts}
}