autoscientist-market-analysis-lenitnes

A LoRA adapter on Qwen/Qwen3.5-9B that turns cryptographic-protocol signal evidence (GitHub commits, releases, diffs, and synthesized cross-repo narratives) into a structured market-analysis verdict — detector-type label, recommended action, confidence, and price direction over {1h, 4h, 24h, 1w} horizons.

Built for the Adaption Labs AutoScientist Challenge Part 2 (Market-Analysis & News category). Trained fully through the AutoScientist API; no manual hyperparameter work.

Results (AutoScientist evaluation)

Metric	Value
Win rate vs base (best checkpoint)	52.56%
Iterations completed	5 / 5
Adaptation quality grade	A
Training rows (adapted dataset)	27,965
Real seed rows (production DB)	1002

Win rate is Adaption's head-to-head metric: the share of held-out evals the adapted model beats the base model on. 52.56% means the adapter wins ~53 of every 100 comparisons — a modest but real, measurable improvement.

Intended use

Research artifact + challenge submission. Given a signal's evidence payload (commit message, release notes, diff patch, or synthesis narrative) it produces a JSON object like:

{"detector_type": "protocol_upgrade", "recommended_action": "review_before_mainnet", "confidence": 0.78, "price_direction": "up", "horizon": "24h"}

Limitations

• Trained on a small real seed (~272 unique signals) expanded platform-side; the lift over base is real but narrow.
• Domain-locked to crypto protocol signals; not a general market analyst.
• Not trading or investment advice. Understands nothing about positions, sizing, or your objectives.

Training provenance

• Base: Qwen/Qwen3.5-9B
• Method: AutoScientist automated loop (5 iterations, LoRA recipe search, platform evaluation keeps best checkpoint)
• Dataset: Papajams/autoscientist-market-analysis-lenitnes-dataset
• Run ID: 4b0ef68e-a233-454b-87d5-203f5c9d401c · Seed dataset ID: c701c50c-7582-4be1-9d7c-62618a001738
• Resume/reproduce: https://github.com/sneldao/lenitnes (docs/AUTOSCIENTIST.md)