VietSurya-v1

Vietnamese document OCR adaptation of Surya OCR 2, released by ANZ Innovation.

VietSurya-v1 is a lightweight LoRA adapter for datalab-to/surya-ocr-2. It packages the Vietnamese adaptation weights together with compatible tokenizer and processor assets. Load the adapter alongside the base model to use Surya's native document OCR workflow.

Model at a glance

Specification	Value
Primary language	Vietnamese
Task	Document image-to-text
Base model	datalab-to/surya-ocr-2
Base architecture	Qwen3.5-style multimodal architecture
Release format	PEFT LoRA adapter · Safetensors
Adapter parameters	18,284,544 · 18.28M
Adapter weight size	73.18 MB · 69.79 MiB
Stored weight dtype	FP32
Training precision	BF16
LoRA rank / alpha	32 / 64
LoRA dropout	0.05
Adapted components	Selected language-model linear-attention and MLP projections
Spatial modules	Frozen during adaptation
Base weights included	No — download the base model separately

Parameter count, storage dtype and file size were verified directly from the published adapter_model.safetensors. The adapter size excludes the base model and tokenizer/processor files.

Checkpoint and validation

Reported training measurement	Value
Selected checkpoint step	22,925
Completed epochs	1
Validation examples	512 · stratified
Selected validation loss	0.04034
Maximum training sequence length	8,192 tokens
Checkpoint selection	Best saved validation-loss candidate

Validation loss measures the training objective; it is not OCR accuracy, CER or WER. This model card does not report CER/WER for this adapter, and the validation loss alone does not establish an improvement over the base model. The training corpus combines Vietnamese document sources; dataset contents are not distributed here.

Usage

The repository contains an adapter, so both the base model and adapter must be accessible to your Hugging Face account. Use a Transformers version that supports the base model's Qwen3.5 architecture, together with compatible PEFT and PyTorch versions.

from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoProcessor

base_id = "datalab-to/surya-ocr-2"
adapter_id = "ANZ-Innovation/VietSurya-v1"

processor = AutoProcessor.from_pretrained(adapter_id)
base_model = AutoModelForImageTextToText.from_pretrained(
    base_id,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

Use Surya OCR 2's documented prompts and processor conventions for generation. Keep the native output format when integrating the model into document-processing workflows. Evaluate OCR text and document structure separately on representative inputs; no exact-layout reproduction guarantee is claimed by this release.

Repository files

File	Purpose
adapter_model.safetensors	Vietnamese LoRA adaptation weights
adapter_config.json	Adapter loading configuration
tokenizer.json, tokenizer_config.json	Compatible tokenizer assets
processor_config.json	Image/text processor configuration
chat_template.jinja	Native prompt formatting
LICENSE	Inherited model license

Training scripts, optimization code, private documents and training datasets are not included in this repository.

Attribution and license

Developed by ANZ Innovation, based on Datalab's Surya OCR 2.

This derivative adapter is distributed under the base model's modified AI Pubs Open Rail-M license, included in LICENSE. It carries use-based restrictions, attribution and share-alike terms. Read the license before using or redistributing the adapter. No endorsement by Datalab is implied.