avraham123/tale-me-dictabert-menaked-browser

DictaBERT Menaked · browser INT8

Independent ONNX conversion of dicta-il/dictabert-large-char-menaked for local Hebrew diacritization in a web browser with ONNX Runtime Web/WASM.

Artifact

File	Bytes	SHA-256
dictabert-menaked.int8.onnx	317,980,253	c97d84ffa008cff1af5fab1b38886ca6f1a120a32d3abf7adf3043b6d21817f1

The model has three dynamic int64 inputs:

• input_ids [batch, sequence]
• attention_mask [batch, sequence]
• token_type_ids [batch, sequence]

It returns nikud_logits [batch, sequence, 29] and shin_logits [batch, sequence, 2] as float32.

The original tokenizer is intentionally not modified. Consumers should pin the official tokenizer.json at source revision d311fbf7c403e50b040440e4859ac78064d025d0 (SHA-256 8e62e3b46c924e14fc32c749ef8944c311411ce9c4dc01c5b606953a169140ba).

Provenance and conversion

• Source model: dicta-il/dictabert-large-char-menaked
• Source revision: d311fbf7c403e50b040440e4859ac78064d025d0
• Source model.safetensors SHA-256: c2927643c61f408c7d5ff1652b605b322ed896fa07ded344bd508a02b76bf50e
• ONNX opset: 17
• Quantization: dynamic signed INT8, per-channel, reduced range
• Export environment: Python 3.12.14, PyTorch 2.2.2, Transformers 4.42.4, ONNX 1.17.0, ONNX Runtime 1.22.1

conversion-manifest.json records the immutable inputs, output hash, model contract, and parity results. export_dicta.py is the conversion and verification program; its environment is pinned in requirements.txt.

Verification

• PyTorch and FP32 ONNX logits matched within rtol=1e-3, atol=1e-4.
• FP32 and INT8 selected the same nikud and shin/sin classes for every relevant Hebrew character in the 12-sentence parity corpus.
• All 12 rendered diacritized sentences were byte-for-byte identical between FP32 and INT8.
• ONNX checker passed and the final session exposed exactly the documented three inputs and two outputs.

These checks establish conversion parity for the recorded corpus; they do not replace the upstream model's broader quality evaluation.

License and attribution

DictaBERT-large-char-menaked © Dicta: The Israel Center for Text Analysis, licensed under Creative Commons Attribution 4.0 International. Converted to ONNX and quantized by Tale Me; modifications were made. Dicta does not endorse Tale Me.

The conversion script is separately available under the MIT license in LICENSE-CODE. The model artifact remains subject to CC BY 4.0.