docs: model card

README.md

@@ -2,7 +2,7 @@
 library_name: ctranslate2
 pipeline_tag: translation
 license: apache-2.0
-base_model: protonx-models/protonx-legal-tc
+base_model: protonx-models/distilled-protonx-legal-tc
 language:
   - vi
 tags:
@@ -12,21 +12,20 @@ tags:
   - vietnamese
   - ocr-correction
   - post-ocr
-  - distillation
 ---
 
-# Distilled protonx-legal-tc — CTranslate2 (Vietnamese OCR text correction)
+# distilled-protonx-legal-tc — CTranslate2 build (Vietnamese OCR text correction)
 
-Distilled + CTranslate2-converted version of [`protonx-models/protonx-legal-tc`](https://huggingface.co/protonx-models/protonx-legal-tc) for fast CPU OCR text correction on Vietnamese administrative documents.
+CTranslate2-converted version of [`protonx-models/distilled-protonx-legal-tc`](https://huggingface.co/protonx-models/distilled-protonx-legal-tc), optimised for fast CPU OCR text correction on Vietnamese administrative documents.
 
-The original protonx-legal-tc is a seq2seq model trained for legal-text correction. This distilled CT2 build is optimized for CPU inference and used by the [ScanIndex](https://github.com/welcomyou/scanindex) pipeline as the correction stage between OCR and PDF/DOCX export.
+The upstream `distilled-protonx-legal-tc` is a smaller student distilled from `protonx-legal-tc`. This repo only does the CT2 conversion + INT8 quantization — no further training. Used by the [ScanIndex](https://github.com/welcomyou/scanindex) pipeline as the correction stage between OCR and PDF/DOCX export.
 
-## Performance (ScanIndex internal benchmark)
+## Performance (ScanIndex internal benchmark, 13-page Vietnamese admin doc)
 
-| Variant | Time / 13 pages | Accuracy |
+| Variant | Time | Accuracy |
 |---|---|---|
-| protonx-legal-tc (CT2 OPTIMIZE, beam=1) | 14.5s | 99.561% |
-| **This distilled CT2 (beam=1)**         | **8.3s** | **99.550%** |
+| `protonx-legal-tc` — CT2 OPTIMIZE, beam=1 | 14.5s | 99.561% |
+| **This repo** (`distilled-protonx-legal-tc` CT2 INT8, beam=1) | **8.3s** | **99.550%** |
 
 42% faster, 0.011 pp accuracy drop — recommended trade-off for CPU.