flamiinngo/market-news-qa

Market News QA — Market-Analysis Instruction Dataset

Concise question-and-answer pairs for market analysis and financial news interpretation: classifying news by market area, reading sentiment, identifying who a story matters to, and answering forward-looking questions from earnings calls.

Built for the Adaption Labs AutoScientist Challenge (Market-Analysis & News category).

Task mix

Five task types, so a model trained on this learns a range of market-analysis skills rather than a single classification behaviour.

Design notes

Answers are short by design. Median 16 words. In head-to-head evaluation, concise factual answers consistently outperformed long explanatory ones.

Answers are grounded in their own headline. Each names the specific subject — "Microsoft here falls under Analyst Update…", "On TOPS, this reads as bullish…" — rather than reciting a generic category description. That is why 85% of answers are distinct despite a fixed set of categories.

Earnings rows are analyst-shaped, not bookkeeping. They ask about outlook, guidance, drivers, and demand rather than "what was revenue in Q3". Pure metric lookups were filtered out because they classify as corporate finance, not market analysis.

Usage

from datasets import load_dataset

ds = load_dataset("flamiinngo/market-news-qa", split="train")
print(ds[0])

Sources & Attribution

Released under CC-BY-4.0, the more restrictive of the upstream licences.

How the earnings rows were filtered

The upstream earnings corpus needed heavy cleaning. From 860,164 rows:

• —292,002 removed as refusals — 34% of the corpus answers "I do not know" or "not mentioned in the transcript" in some phrasing. Training on these teaches a model to decline.
• —207,135 removed for naming no subject — questions asking about "the company's revenue" without saying which company. These teach a model to invent figures for unidentified entities.
• —Further filters removed clinical/pharma content, transcript-dependent answers, and near-duplicate answer structures.
• —Capped at 3 rows per ticker across 880 companies.

Limitations

• —Category labels carry source noise. The upstream annotations are imperfect; some headlines are labelled debatably.
• —Answer explanations are category-level. Each answer names its own subject, but the reasoning clause is drawn from a fixed set of explanations per category. The classification is ground truth; the surrounding analysis is not a per-headline human judgment.
• —Headlines are from Twitter/X and skew toward US markets, 2020–2022.
• —Earnings answers are synthetic QA over real transcripts, generated by the upstream authors rather than written by analysts.
• —Not investment advice. This is training data for research, not a source of financial guidance.

Trained model

Trained with Adaption Labs' AutoScientist on Mixtral-8x7B-Instruct-v0.1 + LoRA.

Disclaimer

For research and informational purposes only. Nothing here is investment advice.