blocks); the stored text is chandra's official markdown export (BatchOutputItem.markdown = chandra.output.parse_markdown with the CLI defaults include_images=True, include_headers_footers=False, i.e. Page-Header/Page-Footer blocks dropped). dtype bfloat16. In this run chandra's client logged 39 repeat-token retry events (the retried pages' final text therefore comes from a sampled, non-greedy re-generation; chandra does not report which pages).",
"software": {
"chandra-ocr": "0.2.0",
"vllm": "0.32.1rc1.dev613+g8977b13a1",
"torch (server)": "2.14.0+cu132",
"transformers (server)": "5.20.2",
"openai (client)": "3.39.0",
"pillow": "13.0.0"
},
"prompt": "OCR this image to HTML, arranged as layout blocks. Each layout block should be a div with the data-bbox attribute representing the bounding box of the block in x0 y0 x1 y1 format. Bboxes are normalized 0-1000. The data-label attribute is the label for the block.\n\nUse the following labels:\n- Caption\n- Footnote\n- Equation-Block\n- List-Group\n- Page-Header\n- Page-Footer\n- Image\n- Section-Header\n- Table\n- Text\n- Complex-Block\n- Code-Block\n- Form\n- Table-Of-Contents\n- Figure\n- Chemical-Block\n- Diagram\n- Bibliography\n- Blank-Page\n\nOnly use these tags ['math', 'br', 'i', 'b', 'u', 'del', 'sup', 'sub', 'table', 'tr', 'td', 'p', 'th', 'div', 'pre', 'h1', 'h2', 'h3', 'h4', 'h5', 'ul', 'ol', 'li', 'input', 'a', 'span', 'img', 'hr', 'tbody', 'small', 'caption', 'strong', 'thead', 'big', 'code', 'chem'], and these attributes ['class', 'colspan', 'rowspan', 'display', 'checked', 'type', 'border', 'value', 'style', 'href', 'alt', 'align', 'data-bbox', 'data-label'].\n\nGuidelines:\n* Inline math: Surround math with tags. Math expressions should be rendered in KaTeX-compatible LaTeX. Use display for block math.\n* Tables: Use colspan and rowspan attributes to match table structure.\n* Formatting: Maintain consistent formatting with the image, including spacing, indentation, subscripts/superscripts, and special characters.\n* Images: Include a description of any images in the alt attribute of an ![]()
tag. Do not fill out the src property. Describe in detail inside the div tag. Also convert charts to high fidelity data, and convert diagrams to mermaid.\n* Forms: Mark checkboxes and radio buttons properly.\n* Text: join lines together properly into paragraphs using
tags for line breaks within paragraphs, but only when absolutely necessary to maintain meaning.\n* Chemistry: Use... tags for chemical formulas with reactive SMILES.\n* Lists: Preserve indents and proper list markers.\n* Use the simplest possible HTML structure that accurately represents the content of the block.\n* Make sure the text is accurate and easy for a human to read and interpret. Reading order should be correct and natural.",
"dtype": "bfloat16",
"hardware": "1 × RTX PRO 6000 Blackwell (96 GB)",
"failures": 0,
"seconds": 712.2,
"all": {
"crops": 767,
"cer": 0.02389145289388281,
"median_crop_cer": 0.014124293785310734,
"edits": 11720,
"reference_characters": 490552,
"turkish_letter_rate": {
"edits": 560,
"reference_letters": 47631,
"rate": 0.011757048980705842
},
"turkish_letter_errors": {
"ı/i/İ/I": {
"errors": 2838,
"reference_letters": 63419
},
"ş": {
"errors": 42,
"reference_letters": 7354
},
"ğ": {
"errors": 82,
"reference_letters": 4147
},
"ç": {
"errors": 21,
"reference_letters": 4827
},
"ö": {
"errors": 66,
"reference_letters": 3547
},
"ü": {
"errors": 38,
"reference_letters": 8018
}
},
"dotless_dotted": 92,
"diacritic_dropped": 42,
"outputs": 767,
"marked_in_text_sent": {
"underlined": {
"spans": 1717,
"marked": 0
},
"numbered": {
"markers": 2862,
"marked": 2528
}
},
"marked_in_raw_output": {
"underlined": {
"spans": 1717,
"marked": 1563
},
"numbered": {
"markers": 2862,
"marked": 2528
}
}
},
"verified": {
"crops": 378,
"cer": 0.04686744317265839,
"median_crop_cer": 0.05211726384364821,
"edits": 9303,
"reference_characters": 198496,
"turkish_letter_rate": {
"edits": 282,
"reference_letters": 19570,
"rate": 0.014409810935104753
},
"turkish_letter_errors": {
"ı/i/İ/I": {
"errors": 2550,
"reference_letters": 26340
},
"ş": {
"errors": 19,
"reference_letters": 3028
},
"ğ": {
"errors": 33,
"reference_letters": 1664
},
"ç": {
"errors": 11,
"reference_letters": 2106
},
"ö": {
"errors": 59,
"reference_letters": 1593
},
"ü": {
"errors": 16,
"reference_letters": 3167
}
},
"dotless_dotted": 31,
"diacritic_dropped": 18,
"outputs": 378,
"marked_in_text_sent": {
"underlined": {
"spans": 1558,
"marked": 0
},
"numbered": {
"markers": 1264,
"marked": 931
}
},
"marked_in_raw_output": {
"underlined": {
"spans": 1558,
"marked": 1412
},
"numbered": {
"markers": 1264,
"marked": 931
}
}
}
},
"dots-mocr": {
"model": "dots-studio/dots.mocr",
"revision": "e539fbb52280393adc081b289ec597430a0f9031",
"method": "dots_mocr parser (github.com/rednote-hilab/dots.mocr @ 23f3e5612fb8, installed editable) official vLLM path: DotsMOCRParser.parse_file(, prompt_mode='prompt_layout_all_en', fitz_preprocess=True), i.e. `python3 dots_mocr/parser.py ` with the CLI defaults (prompt_layout_all_en, top_p 1.0, max_completion_tokens 16384, dpi 200 fitz re-render of the image, no min/max_pixels override) except temperature 0.0 instead of the CLI's 0.1 for greedy decoding; requests go to a vLLM 0.32.1rc1.dev613+g8977b13a1 OpenAI-compatible server (local nightly venv, native DotsOCRForCausalLM implementation, launched as in the model card: --gpu-memory-utilization 0.9 --chat-template-content-format string --served-model-name model --trust-remote-code; model dtype bfloat16 from config). 64 images in flight. The parser's request function was replaced by an identical copy that also records finish_reason (to detect token-limit hits). The model returns layout JSON (bbox, category, text); the stored text is the parser's own markdown export (.md = format_transformer.layoutjson2md(..., no_page_hf=False): cells in reading order joined by blank lines, Formula as LaTeX, Table as HTML, Picture cells as an inline base64 PNG image link); when the JSON fails to parse, the parser's OutputCleaner fallback text is stored (noted in failures).",
"software": {
"dots_mocr": "1.0 (git 23f3e5612fb8066d4034d5ecfc8f33a9243533eb)",
"vllm": "0.32.1rc1.dev613+g8977b13a1",
"torch (server)": "2.14.0+cu132",
"transformers (server)": "5.20.2",
"transformers (client)": "4.57.6",
"openai (client)": "3.39.0",
"pymupdf": "2.0",
"pillow": "12.3.0"
},
"prompt": "Please output the layout information from the PDF image, including each layout element's bbox, its category, and the corresponding text content within the bbox.\n\n1. Bbox format: [x1, y1, x2, y2]\n\n2. Layout Categories: The possible categories are ['Caption', 'Footnote', 'Formula', 'List-item', 'Page-footer', 'Page-header', 'Picture', 'Section-header', 'Table', 'Text', 'Title'].\n\n3. Text Extraction & Formatting Rules:\n - Picture: For the 'Picture' category, the text field should be omitted.\n - Formula: Format its text as LaTeX.\n - Table: Format its text as HTML.\n - All Others (Text, Title, etc.): Format their text as Markdown.\n\n4. Constraints:\n - The output text must be the original text from the image, with no translation.\n - All layout elements must be sorted according to human reading order.\n\n5. Final Output: The entire output must be a single JSON object.\n",
"dtype": "bfloat16",
"hardware": "1 × RTX PRO 6000 Blackwell (96 GB)",
"failures": 0,
"seconds": 280.4,
"all": {
"crops": 767,
"cer": 0.022599031295357067,
"median_crop_cer": 0.006430868167202572,
"edits": 11086,
"reference_characters": 490552,
"turkish_letter_rate": {
"edits": 285,
"reference_letters": 47631,
"rate": 0.005983498141966366
},
"turkish_letter_errors": {
"ı/i/İ/I": {
"errors": 2683,
"reference_letters": 63419
},
"ş": {
"errors": 45,
"reference_letters": 7354
},
"ğ": {
"errors": 45,
"reference_letters": 4147
},
"ç": {
"errors": 19,
"reference_letters": 4827
},
"ö": {
"errors": 10,
"reference_letters": 3547
},
"ü": {
"errors": 44,
"reference_letters": 8018
}
},
"dotless_dotted": 24,
"diacritic_dropped": 10,
"outputs": 767,
"marked_in_text_sent": {
"underlined": {
"spans": 1717,
"marked": 277
},
"numbered": {
"markers": 2862,
"marked": 2262
}
},
"marked_in_raw_output": {
"underlined": {
"spans": 1717,
"marked": 277
},
"numbered": {
"markers": 2862,
"marked": 2130
}
}
},
"verified": {
"crops": 378,
"cer": 0.042816983717556024,
"median_crop_cer": 0.04716981132075472,
"edits": 8499,
"reference_characters": 198496,
"turkish_letter_rate": {
"edits": 139,
"reference_letters": 19570,
"rate": 0.007102708226877874
},
"turkish_letter_errors": {
"ı/i/İ/I": {
"errors": 2371,
"reference_letters": 26340
},
"ş": {
"errors": 26,
"reference_letters": 3028
},
"ğ": {
"errors": 21,
"reference_letters": 1664
},
"ç": {
"errors": 12,
"reference_letters": 2106
},
"ö": {
"errors": 4,
"reference_letters": 1593
},
"ü": {
"errors": 19,
"reference_letters": 3167
}
},
"dotless_dotted": 11,
"diacritic_dropped": 3,
"outputs": 378,
"marked_in_text_sent": {
"underlined": {
"spans": 1558,
"marked": 247
},
"numbered": {
"markers": 1264,
"marked": 682
}
},
"marked_in_raw_output": {
"underlined": {
"spans": 1558,
"marked": 247
},
"numbered": {
"markers": 1264,
"marked": 558
}
}
}
},
"surya-ocr-2": {
"model": "datalab-to/surya-ocr-2",
"revision": "3b3d4cdf88d6928b0acdc75181b13206ea67c4a3",
"method": "surya-ocr 0.22.1 official full-page OCR path (the code behind `surya_ocr `: surya.input.load.load_from_folder, then RecognitionPredictor(SuryaInferenceManager())(images, full_page=True)) against a vLLM 0.20.1 OpenAI-compatible server (vllm==0.20.1 from PyPI in an isolated venv on 1 GPU, the version of surya's vllm/vllm-openai:v0.20.1 docker image, started with the same arguments surya's VllmBackend gives that image for VLLM_GPU_TYPE=h100: --dtype bfloat16 --max-model-len 18000 --max-num-seqs 104 --max-num-batched-tokens 16384 --enable-prefix-caching --mm-processor-kwargs {min_pixels: 3136, max_pixels: 6291456} --speculative-config {method: mtp, num_speculative_tokens: 2}; docker itself was not used). The server venv pins transformers 5.17.0 (surya requires >=5.12.1): transformers >=5.18 loads this checkpoint's char-level tokenizer as the byte-level Qwen3_5Tokenizer, which corrupted every prompt and turned non-ASCII output letters (ı ğ ç ö ü ...) into U+FFFD/control bytes in a first run with vLLM nightly + transformers 5.20 (discarded). Attached via SURYA_INFERENCE_BACKEND=vllm, SURYA_INFERENCE_URL; client concurrency 96 (surya default for that sizing). Each crop is sent once as PNG (after surya's scale_to_fit) with the default full-page prompt HIGH_ACCURACY_BBOX_PROMPT; decoding: temperature 0, top_p 0.1, max_tokens 12288 (surya defaults; SURYA_FULLPAGE_REGEN off), surya's client retries a request that ends in a repeat-token loop at temperature min(0.2*(n+1), 0.8), top_p 0.95 (max 3 retries), and a page whose full-page HTML fails to parse or loops falls back to surya's layout + per-block OCR. The model emits HTML layout blocks (
...
tags. Usetags for line breaks within paragraphs, but only when absolutely necessary to maintain meaning.\n* Chemistry: Use
); surya parses them into PageOCRResult.blocks in reading order with per-block inner HTML. Surya has no markdown/text export of its own (that lives in marker), so the stored text is the non-skipped blocks' `html` (surya's official per-block output, math in