DharmaOCR uses DPO to cut text degeneration rate by 59.4%
Original title:Direct Preference Optimization Beyond Chatbots
AI summary
DharmaOCR, a structured OCR model released in April, used Direct Preference Optimization (DPO) as a second training stage after supervised fine-tuning (SFT) in Brazilian Portuguese document extraction tasks. SFT cannot directly penalize text degeneration (repetition loops), whereas DPO uses the model's own failed outputs (degeneration loops) as negative samples for preference training, reducing text degeneration rates by an average of 59.4% across all tested model families, with a maximum of 87.6% (e.g., Nanonets-OCR2-3B from 1.61% to 0.20%). Traditional DPO is mostly used for chat alignment; this work extends it to objective OCR tasks, demonstrating that DPO can specifically fix particular failure modes.
Key Takeaways
- DharmaOCR,a structured OCR model released in April,used Direct Preference Optimization (DPO) as a…
- SFT cannot directly penalize text degeneration (repetition loops),whereas DPO uses the model's own…
- 4% across all tested model families,with a maximum of 87.
- ,Nanonets-OCR2-3B from 1.
- Traditional DPO is mostly used for chat alignment
Quick Facts
- Published
- Updated
- Source
- Hugging Face 社区博客(混合发现)
- Category
- Tutorials & Tips
- AI score
- 55
DPO 不只能对齐,这篇直接用模型自身的垃圾输出当反例,把 OCR 的重复错误平均降了 59%。思路不复杂,但告诉做结构化生成的工程师:训练数据里最该保留的就是模型犯的错。
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: DharmaOCR uses DPO to cut text degeneration rate by 59.4%. June 3, 2026. https://new.nat6.net/en/news/cmpy47xxi01kislax812354vk/