adelpro/atm-v2-int8-onnx

Arabic Triplet Matryoshka V2 (ATM-V2) — INT8 ONNX (Native)

This is an INT8 dynamic-quantized ONNX export of Omartificial-Intelligence-Space/Arabic-Triplet-Matryoshka-V2, the state-of-the-art Arabic embedding model (ATM2). It is the runtime model bundled on-device by the open-mushaf Quran reader app for its "بالذكاء الاصطناعي" (AI) semantic-search tab.

The model maps Arabic sentences and verses to a 768-dimensional dense vector space for semantic search, similarity, IR, and clustering. It is fine-tuned from aubmindlab/bert-base-arabertv02 and described in the paper GATE: General Arabic Text Embedding for Enhanced Semantic Textual Similarity with Hybrid Loss Training.

Model Description

Arabic-Triplet-Matryoshka-V2-Model is a state-of-the-art Arabic language embedding model based on the sentence-transformers framework and specifically designed to capture the rich semantic nuances of Arabic text.

Attribute	Value
Base model	aubmindlab/bert-base-arabertv02
Embedding dimension	768
Parameters	135M
Pretrained model format	fp32 safetensors
This export	INT8 dynamic-quantized ONNX (~135 MB, graph opset 17)
License	Apache-2.0

Key Features

• State-of-the-Art Performance: Achieved 0.85 on STS17 and 0.64 on STS22.v2 with an average score of 74.5 on Arabic semantic similarity.
• MatryoshkaLoss Training: Nested embeddings at multiple resolutions.
• Full Arabic Support: Handles the complexity and morphological richness of Arabic.
• Small, offline-friendly: ~135 MB INT8 matches a swap-light-mobile budget and runs fully on-device with onnxruntime-react-native.

Training Details

• Loss functions: MatryoshkaLoss + MultipleNegativesRankingLoss.
• Dataset: akhooli/arabic-triplets-1m-curated-sims-len (1M samples).
• Epochs: 3 · Final loss: 0.718 · Embedding dim: 768.

Performance

From the original model card (Arabic semantic similarity benchmarks):

Model	Dim	Params	STS17	STS22-v2	Avg
ATM-V2 (this repo base)	768	135M	85	64	75

See the upstream Omartificial-Intelligence-Space/Arabic-Triplet-Matryoshka-V2 card for the full comparison table.

Use Cases

• Semantic textual similarity, document retrieval, question answering, clustering, and classification of Arabic text.
• On-device semantic search (as used by open-mushaf).

File layout (native, flat directory)

This repo hosts the native flat layout consumed by the React Native app via expo-file-system + onnxruntime-react-native:

model.int8.onnx              (~135 MB, INT8 dynamic-quantized)
tokenizer.json
tokenizer_config.json
special_tokens_map.json

These 4 files are streamed to the app cache on first AI-search use and reused offline afterwards. A companion web layout lives in adelpro/atm-v2-web.

Usage (onnxruntime-react-native)

import { InferenceSession, Tensor } from 'onnxruntime-react-native';

const session = await InferenceSession.create(modelPath, {
  executionProviders: ['cpu'],
  graphOptimizationLevel: 'all',
});

// Tokenize with the matching tokenizer.json, build input_ids + attention_mask.
const results = await session.run({ input_ids, attention_mask });

// Mean-pool the last hidden state over non-pad tokens, then L2-normalize
// to get the 768-dim embedding — identical to what the web runtime produces,
// so the two platforms share one vector space.
const embedding = meanPoolL2Normalize(results);

Note: the vector index built for this model uses mean pooling + L2 normalization with no query:/passage: prefixes (the upstream model card shows plain encode(), not an E5-style prefix convention). Keep query and stored passages in the same unprefixed space.

Limitations

• May not perform optimally on highly technical / domain-specific Arabic that was underrepresented in training.
• A one-word short query (e.g. a proper noun) embeds weakly relative to long passages; semantically-linked epithets may not surface at high cosine.
• Int8 dynamic quantization introduces small weight-precision loss compared to the fp32 original.

Ethical Considerations

Intended for research and applications that benefit Arabic language processing. Biases that may exist in the training data should be considered.

Citation

Cite the upstream paper when referencing this model's methodology:

@article{nacar2025gate,
  title={GATE: General Arabic Text Embedding for Enhanced Semantic Textual Similarity with Matryoshka Representation Learning and Hybrid Loss Training},
  author={Nacar, Omer and Koubaa, Anis and Sibaee, Serry and Al-Habashi, Yasser and Ammar, Adel and Boulila, Wadii},
  journal={arXiv preprint arXiv:2505.24581},
  year={2025}
}

Acknowledgements

Built on aubmindlab/bert-base-arabertv02 and the akhooli/arabic-triplets-1m-curated-sims-len dataset.

---

Derived from Omartificial-Intelligence-Space/Arabic-Triplet-Matryoshka-V2 (Apache-2.0). This ONNX export is the on-device runtime model for the open-mushaf-native Quran reader.