mattboll/mdenseon-onnx

mDenseOn — ONNX export used by Crible

This repository hosts the ONNX export of lightonai/mDenseOn that the integration tests of Crible run. It is not an official LightOn model.

Origin

• Base model: lightonai/mDenseOn
• Upstream revision: a5fdb000f7a21da96c3bddde3a782ef777316df3
• License: Apache-2.0. The upstream repository declares Apache-2.0 in its model card and contains no separate LICENSE or NOTICE file; the licence text is provided here in LICENSE.

Modifications

The files were produced by Crible from the pinned revision above:

• fp32 ONNX export by torch.onnx.export, under transformers 5.14.1;
• INT8 dynamic quantization by onnxruntime.quantization;
• export script: tools/export_onnx.py of the Crible repository, https://github.com/mattboll514/crible.

Encoding contract

The graph returns the last hidden state (last_hidden_state) of the transformer. It applies neither the prompt nor the pooling, which Crible does itself:

• prompts: query: for queries and document: for documents, trailing space included;
• pooling: CLS (position 0 of the output);
• normalization: L2, after pooling;
• dimensions: 768;
• max_tokens: 8192, prompt included;
• sequence length and batch size are dynamic axes.

These values are the ones declared in model.json, which Crible reads.

Files

File	Contents	BLAKE3
model.onnx	fp32 graph, weights inline (1 227 955 436 bytes)	7b6eb68099dbeac6826e282dd575aec766dfbe1bedac143bae094b12f82094a2
model.int8.onnx	INT8 dynamic-quantized graph (307 302 578 bytes)	0534b4ee028cfc4a6a7ca9dcca7a7c908110372033dab33a4f75fbe37d5941a5
tokenizer.json	Hugging Face fast tokenizer, copied as is (34 363 287 bytes)	8a801317cfb1eb744734392dc0aeef36d6d39660a0a884f06a0338e9bb5357db
model.json	Crible model description: revision, prompts, pooling, limits, file digests	19667080a2d779857691186ad1c6199ec72aac0b1fe6fc0ed07f06b6116c737a

The four files are byte-exact copies of the export measured for Crible.

Fidelity

The export is checked against the vectors Crible already produced for Legifacile, and against PyTorch, in the MES-04 campaign (docs/mesures/MES-04-parite-encodeur.md of the Crible repository):

• PyTorch versus ONNX fp32, on 100 deterministic French texts: minimum cosine 0.9999999998854414, mean 0.9999999999971361;
• ONNX fp32 with the document prompt versus the production vectors, on 600 passages of the Legifacile sample: minimum cosine 0.9999603983011769, p1 0.9999935038784599, median 0.9999983369936447. The acceptance threshold of Crible is 0.999.

The test at 1e-4 in Crible is a guard rail; MES-04 is the fidelity measurement.

License

Apache-2.0. See LICENSE for the full text. Any redistribution of these files must keep that licence and this model card.