|
|
@@ -45,8 +45,26 @@
|
|
|
# Changelog
|
|
|
|
|
|
- 2025/09/19 2.5.0 Released
|
|
|
- - vlm update to 2509-2.5 version
|
|
|
|
|
|
+We are officially releasing MinerU2.5, currently the most powerful multimodal large model for document parsing.
|
|
|
+With only 1.2B parameters, MinerU2.5's accuracy on the OmniDocBench benchmark comprehensively surpasses top-tier multimodal models like Gemini 2.5 Pro, GPT-4o, and Qwen2.5-VL-72B. It also significantly outperforms leading specialized models such as dots.ocr, MonkeyOCR, and PP-StructureV3.
|
|
|
+The model has been released on [huggingface](https://huggingface.co/opendatalab/MinerU2.5-2509-1.2B) and [ModelScope](https://modelscope.cn/models/opendatalab/MinerU2.5-2509-1.2B) platforms. Welcome to download and use!
|
|
|
+- Core Highlights:
|
|
|
+ - SOTA Performance with Extreme Efficiency: As a 1.2B model, it achieves State-of-the-Art (SOTA) results that exceed models in the 10B and 100B+ classes, redefining the performance-per-parameter standard in document AI.
|
|
|
+ - Advanced Architecture for Across-the-Board Leadership: By combining a two-stage inference pipeline (decoupling layout analysis from content recognition) with a native high-resolution architecture, it achieves SOTA performance across five key areas: layout analysis, text recognition, formula recognition, table recognition, and reading order.
|
|
|
+- Key Capability Enhancements:
|
|
|
+ - Layout Detection: Delivers more complete results by accurately covering non-body content like headers, footers, and page numbers. It also provides more precise element localization and natural format reconstruction for lists and references.
|
|
|
+ - Table Parsing: Drastically improves parsing for challenging cases, including rotated tables, borderless/semi-structured tables, and long/complex tables.
|
|
|
+ - Formula Recognition: Significantly boosts accuracy for complex, long-form, and hybrid Chinese-English formulas, greatly enhancing the parsing capability for mathematical documents.
|
|
|
+
|
|
|
+Additionally, with the release of vlm 2.5, we have made some adjustments to the repository:
|
|
|
+- The vlm backend has been upgraded to version 2.5, supporting the MinerU2.5 model and no longer compatible with the MinerU2.0-2505-0.9B model. The last version supporting the 2.0 model is mineru-2.2.2.
|
|
|
+- VLM inference-related code has been moved to [mineru_vl_utils](https://github.com/opendatalab/mineru-vl-utils), reducing coupling with the main mineru repository and facilitating independent iteration in the future.
|
|
|
+- The vlm accelerated inference framework has been switched from `sglang` to `vllm`, achieving full compatibility with the vllm ecosystem, allowing users to use the MinerU2.5 model and accelerated inference on any platform that supports the vllm framework.
|
|
|
+- Due to major upgrades in the vlm model supporting more layout types, we have made some adjustments to the structure of the parsing intermediate file `middle.json` and result file `content_list.json`. Please refer to the [documentation](https://opendatalab.github.io/MinerU/reference/output_files/) for details.
|
|
|
+
|
|
|
+Other repository optimizations:
|
|
|
+- Removed file extension whitelist validation for input files. When input files are PDF documents or images, there are no longer requirements for file extensions, improving usability.
|
|
|
|
|
|
<details>
|
|
|
<summary>History Log</summary>
|
myhloli