OpenBMB's 2B OCR correctness judgment model (trained from Qwen3.5-2B, with the output head replaced by a binary classification head): It takes as input the original image of a text, formula, or table region and the recognition result provided by an OCR model, and outputs a score between 0 and 1, where a higher score indicates a higher likelihood of correctness. It is used to determine which regions are automatically approved and which are routed for manual review, or to select the best result among outputs from multiple OCR models.
Upvotes 67 · Likes 67 · Downloads 963
Reasons
- Officially claimed to achieve an AUROC (×100) of 93.6 on their proprietary JevOCRBench, surpassing JEV-35B-A3B's 90.1 and approaching Qwen3.5-397B-A17B's 93.8.
- For parsing results from MiniCPM-OCR 1.0, GLM-OCR, PaddleOCR-VL-1.6, and Xiaomi-OCR-0, the AUROC ranges between 91.9 and 93.4 (based on the same batch of 200 samples with manually annotated correctness), and it is not tied to any specific OCR provider.
- It is a companion model to MiniCPM-OCR 1.0, which was included yesterday. It is licensed under Apache-2.0 and has GGUF, AWQ, and GPTQ versions; it was not included in previous issues.
Usage
- Deployment follows the tutorials in GitHub OpenBMB/MiniCPM-OCR: Documentation is available for Docker / local SDK, vLLM, SGLang, llama.cpp, Ollama, and FlagOS. Services with a renderer only require passing the original region image, candidate text, and region category (text / formula / table); services without a renderer require you to render the candidate text into an image and pass it along.
- Usage 1: Score each region, set a threshold based on your business samples, automatically approve regions above the threshold, and route the rest for manual review. Usage 2: Under the same layout segmentation, have several OCR models parse the document, then select the result with the highest score for each region to reconstruct the full text.
- Only when the service returns a verdict of 'scored' does it represent the model's actual score; fixed values returned during pre-checks, rendering, or forward pass failures should not be interpreted as low scores.
Risks
- The evaluation set JevOCRBench is self-built by the official team, and cross-model experiments involve only 200 samples; the official team advises verifying performance yourself when switching parsers or handling new document types.
- It can only select from existing candidates and does not correct recognition errors; its effectiveness is limited by the quality of the candidates themselves. Candidates for the same region must use identical cropping boundaries.
- The cloud API documentation is still a placeholder draft; VRAM requirements are not specified (refer to the FAQ).