pepperocr-1-4b-GGUF

PepperOCR-1-4B is a 4.54-billion-parameter multilingual vision-language model for document parsing, fine-tuned from Qwen3.5-4B (architecture Qwen3_5ForConditionalGeneration) by sionic-ai, designed to convert document images directly into Markdown output and released as a public pre-release evaluation repository with weights distributed in BF16 safetensors (2 shards, ~9.1 GB). The model ships with a fully reproducible MDPBench evaluation runtime locked via uv with stable vLLM 0.24.0, PyTorch 2.11.0+cu130, CUDA 13.0, and Transformers 5.11.0, processing all 17 public MDPBench languages with a CPU-based document-orientation classifier (PP-LCNet_x1_0_doc_ori) preceding inference; on locally reproduced Korean and Thai public-set scores it achieves 92.2 and 83.2 overall accuracy respectively (87.7 KO/TH macro average), broken down further into digital versus photographed document subsets, using a greedy-first decoding strategy with a degeneration guard that retried roughly 10% of samples. As a pre-release evaluation preview, no final license has yet been applied to the fine-tuning contribution itself (though the underlying Qwen3.5-4B base remains Apache-2.0, and the final PepperOCR-1-4B release is planned under Apache-2.0 pending evaluation approval), and public downloadability does not currently grant redistribution or reuse rights.

Model Files

llama.cpp