CritiqueBank-11M: A Large-Scale Dataset of Adversarial Critiques over Generated Trading Rationales

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

We release CritiqueBank-11M, a dataset of 11,674,598 rationale–critique pairs for training models that critique AI-generated trading rationales. Each example pairs a generated trading rationale with an adversarial critique that challenges its assumptions, timing, or evidential basis. The dataset is deliberately asymmetric. The 215 distinct rationales are the complete cross product of 43 tickers and 5 indicator templates — Bollinger Band expansion, Golden Cross, MACD bullish crossover, RSI oversold and Volume spike — while critiques vary across six labelled styles. The corpus therefore provides many critique formulations per input rather than broad input coverage, and suits preference-pair construction, critique-style conditioning, and training a critic to vary its angle of attack. A model trained on it has seen 215 distinct inputs and should not be assumed to generalise to unseen rationale forms without evaluation. Critiques were generated by Google Gemini Flash. No human annotation, adjudication, or inter-annotator agreement statistic exists for this corpus. CritiqueBank-11M was used to train the MiniCrit-7B research checkpoint, released alongside it. This version supersedes the initial deposit of 6 January 2026. It corrects the example count from 11,742,891 to 11,674,598, and withdraws a six-domain distribution, a five-category flaw taxonomy, a flawed/sound split, a benchmark comparison against other models, and production deployment results — none of which the released dataset supports. The stated licence is corrected to match the released artifact. A full table of changes appears at the end of the paper.