PiterQ/mmlw-retrieval-roberta-large-v2-CoreML

mmlw-retrieval-roberta-large-v2 — Core ML (float16)

Core ML conversion of sdadas/mmlw-retrieval-roberta-large-v2 by Sławomir Dadas (OPI PIB), for on-device retrieval (search in recordings and documents) in the AudioTranscriber app (iPhone and Mac; Neural Engine / GPU).

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms.

The original model is a derivative subject to the Gemma Terms of Use (training data prepared with a Gemma-based model), so this conversion is distributed under the same terms. Use of this model must comply with the Gemma Terms of Use (copy: GEMMA_TERMS_OF_USE.md) and the Gemma Prohibited Use Policy (copy: GEMMA_PROHIBITED_USE_POLICY.md), which are incorporated by reference. See also NOTICE.

Files

File	Contents
model.mlmodelc/	compiled Core ML model (mlprogram, float16 weights)
meta.json	dimensions, max tokens, input shapes, query prefix, special token ids
tokenizer.json, tokenizer_config.json, special_tokens_map.json, config.json	original tokenizer and config
reference.json	reference token ids and PyTorch (fp32) embeddings for 5 test sentences (conversion check)

Usage

• Inputs: input_ids and attention_mask (int32, shape [1, n], n ∈ 512), right-padded with the pad token id 1; maximum 512 tokens.
• Output: embedding — float32 [1, 1024], already L2-normalized (cosine = dot product).
• Queries must be prefixed with [query]: (as in the original model card); passages are encoded without a prefix.

Changes made (conversion)

• PyTorch → Core ML mlprogram (coremltools), float16 compute precision, enumerated input shapes;
• CLS pooling and L2 normalization moved into the graph (same as the original Sentence-Transformers configuration);
• attention mask computed explicitly (additive 4D mask), encoder layers unchanged;
• no fine-tuning, no change of weights other than the float16 format.

Conversion script: AudioTranscriber tools/rag_coreml_convert.py (source revision of the original: 13701958e518c47b1f3f91b4d2cdd5afb3eb9331). Quality of the original model (benchmark by the app authors, nDCG@10): PolQA (PL) 0.771, SciFact-PL 0.789, SciFact (EN) 0.786.

Attribution

• Original model: Sławomir Dadas, National Information Processing Institute (OPI PIB) — sdadas/mmlw-retrieval-roberta-large-v2.
• Gemma is a trademark of Google LLC. This repository is not affiliated with or endorsed by Google or OPI PIB.