Reproduce “Jina-OCR-v1: Efficient Document Parsing with Speculative Decoding and Dense Verifiable Rewards” on OpenResearch

About this paper

Alejandro Barón García, Feng Wang, Emilia Garcia Casademont, Han Xiao

An end-to-end document parsing model, Jina-OCR-v1, improves efficiency on low-budget GPUs through speculative decoding while enhancing parsing quality with a reinforcement learning framework using dense verifiable rewards. It achieved an OmniDocBench v1.6 score of 91.14 and a 1.95x speedup on L4 GPUs.

Key insights

• Speculative decoding techniques like FastMTP can yield substantial inference speedups (up to 1.95x on L4 GPUs) for document parsing VLMs, leveraging the strong local structure of their outputs.
• Rich, verifiable reward signals that grant partial credit for structured content are highly effective in guiding reinforcement learning to improve the fidelity of complex document parsing tasks.
• A comprehensive post-training strategy encompassing precise data curation, synthetic data generation, and instruction alignment is critical for developing robust and generalizable document parsers.

Summary by alphaXiv.