Reza2kn/Bina-0.1-Koochik-CoreML

Bina 0.1 CoreML — بینا ۰.۱ برای CoreML

This is the fixed-512 CoreML runtime for Reza2kn/Bina-0.1-Koochik, the merged BF16 Persian OCR baseline.

Precision: BF16 source, mixed CoreML runtime

This is not labeled as a literal BF16 CoreML model. Core ML Tools 9.0 does not expose BF16 as an ML Program compute precision. The parity-qualified build uses:

• vision encoder: CoreML FP32;
• language prefill: CoreML FP16;
• autoregressive decode step: CoreML FP16;
• host token embeddings and M-RoPE tables: FP32.

An all-FP16 attempt was rejected because FP16 vision changed the first greedy token. Keeping vision in FP32 restored exact greedy transcripts while retaining FP16 for the two large language graphs.

Parity evidence

All comparisons use deterministic greedy decoding, EOS token 2, the same BF16 source checkpoint, and the exact Bina transcription prompt:

OCR this image. Return only the exact text visible in the image, preserving Persian, numbers, line breaks, and punctuation. Do not explain.

Gate	Result
Synthetic compiled CoreML canary	22/22 tokens exact through EOS
Two real Persian pages	128/128 checked tokens exact (64-token cap per page)
Real Persian line crop, Python-hosted CoreML	25/25 tokens and text exact through EOS
Same Persian crop, native Swift host	25/25 tokens and text exact through EOS

The native crop transcript is exactly:

گزارش سالانه دیتاک -۱۴۰۲

Receipts are under validation/. The Swift receipt exercises the actual CoreML packages, FP32 host assets, cache updates, baked M-RoPE tables, Fuse tokenizer decoding, and EOS stopping without loading the PyTorch model.

Runtime contract

This release intentionally has one static shape:

• centered white 512 × 512 RGB canvas;
• pixel_values: [1024, 1536], Float32;
• 256 merged image tokens;
• prompt length: 416 tokens;
• M-RoPE delta: -240 (baked into the native decode tables);
• cache length: 1024;
• maximum generated length before the cache boundary: 608 tokens.

The qualified preprocessing path is PIL RGB + LANCZOS aspect-fit into the centered white canvas, followed by the checked-in processor. The Swift generate(pixelValues:) path is parity-qualified. The convenience Swift generate(image:) path uses CoreGraphics and is not claimed to be pixel-identical to PIL.

This is not yet the six-aspect-bucket Bina/Flax runtime and is not a drop-in replacement for every dynamic Surya page shape. Use this release only for the fixed-512 contract above.

Files

• surya_vision_fp32.mlpackage — Bina-specific vision encoder.
• surya_prefill_fp16_seq416_cache1024.mlpackage — language prefill and initial cache.
• surya_decode_step_fp16_cache1024.mlpackage — one autoregressive decode step.
• native_assets/ — FP32 token embedding, prompt embeddings, M-RoPE tables, tokenizer vocabulary, constants, and a synthetic canary tensor.
• native/SuryaCoreMLRuntime/ — Swift Package for macOS 14+ / iOS 17+.
• processor/ — exact tokenizer and processor configuration.
• scripts/ — conversion, fixture preparation, and parity validators.

The internal Surya names reflect the upstream architecture; all three graphs contain the Bina 0.1 fine-tuned weights.

Native smoke test

Clone with Git LFS, then run the release build from the repository root:

git lfs install
git lfs pull

cd native/SuryaCoreMLRuntime
swift run -c release surya-coreml-smoke \
  --model-dir ../.. \
  --pixel-values-fp32 ../../native_assets/canary_pixel_values_fp32.bin \
  --max-tokens 64

Expected token IDs:

2009 2046 2054 2047 2041 2035 2037 1979 136 1072 2020 2047 2052 2033 2044 1979 1982 55 1992 13 13 2

Expected text:

Invoice 123<br/>Total $42.00

To prepare an exact fixed-512 tensor for another image:

python scripts/prepare_bina_coreml_fixtures.py \
  --runtime-script scripts/export_surya_coreml_runtime.py \
  --image /path/to/image.png \
  --output-dir prepared

Then pass prepared/<stem>.pixel_values_fp32.bin to the Swift executable.

Source provenance

• source model: Reza2kn/Bina-0.1-Koochik;
• source revision: 9b5812be94e3e015142da22a4b61b34aad5d2c02;
• source model.safetensors SHA-256: 2193be4ef3d2366438121a15b7a1dea2bb85b24f83145e5a39bfa1f387891ada;
• source config.json SHA-256: e0de22be177070f206106c184d062176fcda591d9114068c42489ffc550488de;
• architecture: Qwen3_5ForConditionalGeneration;
• conversion host: Apple M2, macOS 26.5.1;
• toolchain: Python 3.11.11, PyTorch 2.12.0, Transformers 5.13.1, Core ML Tools 9.0, Xcode 26.5.

Package and native-asset hashes are recorded in coreml_runtime_manifest.json.

خلاصه فارسی

این مخزن نسخهٔ CoreML مدل بینا ۰.۱ است. وزن مبدأ BF16 است، اما چون CoreML خروجی محاسباتی BF16 ارائه نمی‌کند، انکودر تصویر با FP32 و بخش زبانی با FP16 تبدیل شده است. در آزمون واقعی فارسی، خروجی CoreML و مدل اصلی BF16 تا توکن EOS کاملاً یکسان بود؛ همین آزمون با میزبان Swift نیز ۲۵ از ۲۵ توکن و متن دقیقاً یکسان داد. این نسخه فعلاً فقط قرارداد ثابت ۵۱۲×۵۱۲ را پشتیبانی می‌کند.

License

This derivative follows the source model's OpenRAIL license. Review the source and upstream Surya OCR 2 terms before redistribution or deployment.