I released a preview of an oracle-verified reasoning dataset generated by a deterministic Python oracle (no LLM-generated labels).
Key feature: verify-or-fix records with explicit witnesses (counterexamples / better solutions) and bounded reasoning traces (e.g., Dijkstra pop/relax, DP transitions).
Preview: Giuelvy/diamond-logic-miner-preview · Datasets at Hugging Face
Pilot (gated): Giuelvy/diamond-logic-miner-pilot-1m-gold · Datasets at Hugging Face
Happy to share quick stats (task mix, hard-positive rate, length p95) if helpful.