minnb/investment-bias-corpus

Investment Bias Corpus

Summary. A paired corpus for evaluating investment decision-making under behavioral biases.
It contains 1,000 neutral investment samples and 5,000 biased rewrites (anchoring, herding, recency, loss aversion, confirmation) with facts held constant so any change in model behavior can be attributed to narrative framing rather than altered data.

Files

• neutral.jsonl.gz — 1,000 neutral items
• anchoring.jsonl.gz — 1,000 biased items
• herding.jsonl.gz — 1,000 biased items
• recency.jsonl.gz — 1,000 biased items
• loss_aversion.jsonl.gz — 1,000 biased items
• confirmation.jsonl.gz — 1,000 biased items

Each biased row is a rewrite of a neutral row with the same facts and a different narrative frame.

Schema

Each line is one JSON object with the following keys:

• id — unique row ID (e.g., anchoring_0123, neutral_0123)
• parent_id — the neutral item ID that a biased row derives from (e.g., 0123)
• bias — one of: none, anchoring, herding, recency, loss_aversion, confirmation
• context — decision-relevant text (facts, figures, tickers, dates)
• question — explicit yes/no investment question that names the asset/opportunity
• answer — Yes or No (ground truth)
• source — URL or generator tag (e.g., “ChatGPT 5 generated”, “Grok 3 generated”, “Perplexity generated”)
• category — one of: Financial Reports, Financial News, Social Media, LLM generated, Template-Based

Bias Generation & Invariants

Biased rewrites apply canonical behavioral frames while keeping all facts unchanged:

• Hard invariants: numbers, tickers, dates, company names remain identical; paragraph count preserved; total length kept within ±20% of the neutral version; no fabricated facts.
• Framing only: anchoring (overweight prior reference points), herding (crowd/consensus emphasis), recency (overweight newest data), loss aversion (downside/risk emphasis), confirmation (selective highlighting of supportive details).

How to Load

from datasets import load_dataset

ds = load_dataset(
    "minnb/investment-bias-corpus",
    data_files={
        "neutral": "neutral.jsonl.gz",
        "anchoring": "anchoring.jsonl.gz",
        "herding": "herding.jsonl.gz",
        "recency": "recency.jsonl.gz",
        "loss_aversion": "loss_aversion.jsonl.gz",
        "confirmation": "confirmation.jsonl.gz",
    }
)

Intended Use & Limitations

• Intended for research on model robustness, bias detection/mitigation, evaluation, and prompt/decoding strategies in financial contexts.

• Biased texts use narrative framing rather than content alteration, so they may under-approximate real-world discourse intensity.

• Ensure compliance with third-party content rights for any included summaries/excerpts.

Citation

If you use this dataset, please cite:

Minn B. (2025). Investment Bias Corpus (Neutral + 5 Bias Buckets). 
Hugging Face: minnb/investment-bias-corpus

Contact

For questions or issues, please open a discussion on the dataset page or file an issue in the repository.