cemoss17/chess-position-embedder-onnx

Chess position embedder (ONNX)

ONNX export of the 9M-parameter DeepMind Searchless Chess action-value model, used as a FEN → 768-dim position embedding. The action conditioning is held fixed (action=0) so it contributes a constant offset; we mean-pool the trunk hidden state over the 77 FEN positions to produce a position-only embedding.

Side-to-move is mirrored to white via python-chess before tokenization so embeddings live in a canonical frame.

Files

• searchless-9m-mma.onnx — ONNX graph (external data references in .onnx.data)
• searchless-9m-mma.onnx.data — external tensor weights

Both files must be downloaded side-by-side; the .onnx references the .onnx.data by relative path.

Provenance

Ported from JAX weights via web/scripts/onnx/port_weights.py and web/scripts/onnx/export.py. JAX↔ONNX numerical parity was verified to max abs diff 6.1e-05 on a 128-FEN fixture (web/scripts/onnx/test_parity.py).