qwen3-4b-move-intent-v2

Qwen3 4B Move Intent v2

This private development adapter maps one finalized English chess transcript to Move Interpretation v2. The transcript may contain ordinary speech or controlled ASR noise. The model does not see the board and does not produce SAN. A separate resolver must combine the interpretation with the current chess.js position.

Behavior spec

Given one finalized English chess transcript and no board state, the model returns exactly one canonical move-interpretation/v2 compact value that preserves every spoken move constraint, or UNKNOWN when the transcript does not support one coherent interpretation. It emits no additional text or whitespace and never infers legality, SAN, check, mate, or an unspoken constraint.

This is the canonical move-intent-behavior/v2 contract. The model sees no FEN, legal moves, move history, SAN, ASR confidence, transcript alternatives, or partial transcript. UNKNOWN is valid for conflicting, contextual, non-move, or meaning-destroyed text. Malformed output is an adapter failure, not UNKNOWN.

The repository includes the exact human-readable behavior spec and its machine-readable grammar.

Output contract

The model emits one of these compact values:

UNKNOWN
O-O
O-O-O
M|piece|destination|source-square|source-file|source-rank|capture|promotion|special
R|piece|destination|source-square|source-file|source-rank|-|promotion|-

Piece codes are P, N, B, R, Q, and K. x means a spoken capture, ep means en passant, and - means the constraint was not spoken. R records a spoken recapture. The trained newline is a transport frame: inference stops at the first tokenizer token containing that newline, removes the newline, and sends the remaining compact value to the strict parser. It does not trim or repair the value.

The resolver should enumerate legal chess.js moves, filter them using the spoken constraints, and accept only one match. chess.js can then return SAN. Zero matches and multiple matches should fail closed.

Training run

The adapter starts from Qwen/Qwen3-4B-Base at revision 906bfd4b4dc7f14ee4320094d8b41684abff8539. No previous adapter was loaded.

Training used 10,800 fresh Move Intent v2 rows for one epoch: 3,000 authored, 3,000 natural, 3,000 synthetic-ASR, and 1,800 unknown examples. All legacy_relabelled rows were excluded. Heldout data was not packaged or read.

The recipe used NF4 QLoRA with BF16 compute, rank 16, batch size 8, and 1,350 optimizer updates. A fresh process resumed from checkpoint 675. Both segments had finite losses and gradients. Training ran for 899 billed seconds on one NVIDIA A100 80 GB; the estimated provider charge was $0.624 at $2.50 per hour.

Development result

The unchanged adapter was reloaded with the pinned base and evaluated on 128 development rows. The corrected strict replay produced:

The replay used the same row IDs and greedy generation settings as the original verification. It passed the configured development gate.

Status and limits

This is a development pilot, not an official production candidate. The corpus manifest is still marked production_ready: false. Human review, a frozen baseline comparison, and audio-derived ASR evaluation remain open. The sealed heldout split has not been evaluated.

Ten development errors remain. They include promotion phrases such as "make it a rook," one queenside-castling miss, and a few source-constraint mistakes. Board legality is outside this model by design and belongs in the chess.js resolver.

Provenance

• Training request: d9139d244d246f79b278610cd4e61de557746792b304c55b664906ddd36d14d7
• Training job: 6a89df7d7c5c7dd37923505c
• Verification replay job: 6a89e49a73304676c8ec860a
• Dataset split identity: 2414d6b289e3d96fe18392fc5eaba425506c75d6f6265c9453f8aaadb37e4545
• Adapter weights SHA-256: 1d10f6509bce21c058281630a94e84a232088fb72ac57939a47c92fcab94c62d
• Corrected verification SHA-256: 6c851e2924b305fda9b3953a25513d8b5ced88c6c1c582129997d9026efcc665