Papajams/autoscientist-market-analysis-lenitnes-dataset

autoscientist-market-analysis-lenitnes-dataset

The adapted dataset used to fine-tune Papajams/autoscientist-market-analysis-lenitnes for the Adaption Labs AutoScientist Challenge Part 2 (Market-Analysis & News category).

Composition

Seed provenance (real data): the lenitnes production platform monitors 12+ cryptographic protocol repositories around the clock (bitcoin, solana/agave, halo2, and others) and classifies 15 detector types (protocolupgrade, securitycritical_patch, dependency-risk, …). Each signal carries measured market outcomes at four horizons — 1h, 4h, 24h, 1w — so one signal legitimately yields up to four training examples (same evidence, different label and task per horizon).

Split discipline: all split logic operates on signal_id, never on examples, so horizon variants of one signal stay on the same side of any train/eval split.

Direction balance (seed): flat 366 / down 331 / up 305.

Format

Instruction (prompt/completion) JSONL. The completion target is a JSON object with the signal verdict:

{"detector_type": "...", "recommended_action": "...", "confidence": 0.0, "price_direction": "up|down|flat", "horizon": "1h|4h|24h|1w"}

Licensing and provenance note

Some prompt text embeds snippets of third-party open-source code (release diffs) from the monitored repositories. The adapted rows are released under cc-by-4.0; verify upstream repository licenses before reusing embedded code.

Citation

lenitnes team (2026). autoscientist-market-analysis-lenitnes-dataset.
AutoScientist Challenge Part 2 submission, Market-Analysis & News.