rrodolfo0/move-intent-v2-final

Move Intent v2 final

This is the exact submitted transcript-only Move Intent corpus. Given one finalized English chess transcript and no board state, the target is one canonical move-interpretation/v2 compact value or UNKNOWN.

The target is Move Interpretation, not SAN. The model never receives FEN, legal moves, move history, resolver candidates, or SAN. A host-owned Move Resolver checks the interpretation against chess.js legal moves; chess.js derives legality, playable identity, check, mate, and SAN.

Frozen contents

The production identity is 95f7651992d3ec5d1fbd52bbb53b56514997e5a0cf7390ad7c2ad6452c2fc352. The release contains 10,800 fresh training rows, 1,500 development rows, a frozen 300-row own-test panel, and 2,100 remaining holdout rows. The four training selections are nested prefixes of 1,350, 2,700, 5,400, and 10,800 rows. All 1,200 legacy_relabelled source-pool rows are excluded from submitted training selections.

manifest.json records file hashes, row counts, coverage, leakage checks, review results, and provenance. The 160-row review accepted 158 labels and corrected two legacy labels that had dropped a spoken king constraint. It used transcripts and the compact contract only. Controlled noisy-ASR rows are synthetic text corruptions, not evidence from recorded audio.

Files

The JSONL files contain transcript-to-interpretation rows. train.hf.jsonl is the conversational training view. Selection manifests bind every frozen split and curve prefix. SHA256SUMS and content-manifest.json cover the publication bytes added around the production manifest.

Limits

No controlled curve checkpoint met every registered own-test gate. Staff heldout was not supplied. The public own-test labels are research evidence; they are not a substitute for an unseen staff set.