{
"_meta": {
"description": "Machine-readable catalog of the OCR recipes in this directory: script -> model, params, backend, support level, the tested Jobs launch config (the script's [tool.hf-jobs] header), and what the model's own card claims about language coverage. Human-readable language notes: LANGUAGES.md. Language evidence levels, strongest first: per-language-benchmark | count-claim | named-list | multilingual-unspecified | english-only | not-stated | not-applicable. Claims come from model cards (compiled 2026-07-15) - treat as claims, not guarantees.",
"updated": "2026-10-30",
"run_pattern": "hf jobs uv run https://huggingface.co/datasets/uv-scripts/ocr/raw/main/<script> <input-dataset> <output-dataset> (hf CLI >= 1.32 reads image/flavor/secrets from the script's [tool.hf-jobs] header; add --timeout for long runs. Scripts without a `jobs` entry need --flavor/-s HF_TOKEN (and any --image in their docstring).)",
"support_levels": {
"core": "Maintained. Smoke-tested flagless on HF Jobs on 2026-09-23/24 with the pinned launch config in `jobs`.",
"tool": "Maintained companion (judge, layout, text extraction), not a page-OCR recipe.",
"legacy": "Works (smoke-tested 2026-09-23) but superseded or little used; no header yet. See support_note for the alternative.",
"unsupported": "Known broken as of 2026-09-23; kept in the repo for now. See support_note.",
"experimental": "Works in testing on HF Jobs with the launch config in `jobs`, but changes often and may break. See support_note for the stable alternative."
}
},
"tesseract-ocr.py": {
"model_id": null,
"model_name": "Tesseract 5",
"params": null,
"backend": "pytesseract (CPU)",
"task": "ocr",
"output": "plain text",
"license": "apache-2.0",
"languages": {
"evidence": "named-list",
"count": "125 traineddata packs (~100+ languages + script models)",
"list_url": "https://github.com/tesseract-ocr/tessdata_best",
"notes": "Broadest named coverage; script models (Latin, Cyrillic, Devanagari, ...) cover languages without a dedicated pack. Legacy baseline quality."
},
"support": "core",
"jobs": {
"flavor": "cpu-upgrade",
"secrets": [
"HF_TOKEN"
]
}
},
"pp-ocrv6.py": {
"model_id": "PaddlePaddle/PP-OCRv6",
"params": "1.5M-34.5M",
"backend": "paddleocr",
"task": "ocr",
"output": "plain text",
"license": "apache-2.0",
"languages": {
"evidence": "count-claim",
"count": 48
},
"support": "core",
"jobs": {
"flavor": "t4-small",
"secrets": [
"HF_TOKEN"
]
}
},
"falcon-ocr.py": {
"model_id": "tiiuae/Falcon-OCR",
"params": "0.3B",
"backend": "falcon-perception",
"task": "ocr",
"output": "markdown",
"license": "apache-2.0",
"languages": {
"evidence": "not-stated"
},
"support": "legacy",
"support_note": "Falcon-OCR v1; v1.5 work is in PR #110."
},
"smoldocling-ocr.py": {
"model_id": "ds4sd/SmolDocling-256M-preview",
"params": "256M",
"backend": "transformers",
"task": "ocr",
"output": "DocTags",
"languages": {
"evidence": "english-only"
},
"support": "unsupported",
"support_note": "Broken (2026-09-23): the DocTags-to-markdown converter is a stub, so rows are raw DocTags."
},
"surya-ocr.py": {
"model_id": "datalab-to/surya-ocr-2",
"params": "0.65B",
"backend": "vllm",
"image": "vllm/vllm-openai:v0.20.1",
"task": "ocr | layout | table",
"output": "markdown + surya_blocks (per-block HTML, bboxes, reading order)",
"license": "modified OpenRAIL-M",
"languages": {
"evidence": "per-language-benchmark",
"count": 91,
"benchmark_url": "https://github.com/datalab-to/surya/blob/master/static/docs/multilingual.md",
"notes": "38/91 languages score >=90%, 76/91 >=80%; weakest of top-15: Arabic 72.7%, Vietnamese 73.2%."
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.20.1",
"python": "/usr/local/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/site-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
}
},
"glm-ocr.py": {
"model_id": "zai-org/GLM-OCR",
"params": "0.9B",
"backend": "vllm",
"task": "ocr",
"output": "markdown",
"languages": {
"evidence": "named-list",
"count": 8,
"named": [
"zh",
"en",
"fr",
"es",
"ru",
"de",
"ja",
"ko"
],
"notes": "Declared in card metadata only; no per-language evidence in the card body."
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.29.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
},
"image": "vllm/vllm-openai:v0.29.0"
},
"paddleocr-vl.py": {
"model_id": "PaddlePaddle/PaddleOCR-VL",
"params": "0.9B",
"backend": "vllm",
"task": "ocr | table | formula | chart",
"output": "markdown",
"languages": {
"evidence": "count-claim",
"count": 109,
"notes": "Names scripts (Cyrillic, Arabic, Devanagari, Thai); claims handwriting + historical documents; no per-language numbers."
},
"support": "legacy",
"support_note": "PaddleOCR-VL 1.0; use paddleocr-vl-1.6.py."
},
"paddleocr-vl-1.5.py": {
"model_id": "PaddlePaddle/PaddleOCR-VL-1.5",
"params": "0.9B",
"backend": "transformers",
"task": "ocr (6 modes)",
"output": "markdown",
"languages": {
"evidence": "count-claim",
"count": "109+ (adds Tibetan, Bengali)",
"notes": "Claims improved rare-character and ancient-text recognition."
},
"support": "unsupported",
"support_note": "Broken (2026-09-23): every row is an [OCR ERROR] ('min_pixels' attribute removed in transformers 5). Use paddleocr-vl-1.6.py."
},
"paddleocr-vl-1.6.py": {
"model_id": "PaddlePaddle/PaddleOCR-VL-1.6",
"params": "0.9B",
"backend": "vllm",
"task": "ocr",
"output": "markdown",
"languages": {
"evidence": "count-claim",
"count": "109+ (inherited from 1.5, not restated)",
"notes": "Upgrade-focused card; language-specific claims are Chinese ancient documents and rare characters."
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.29.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
},
"image": "vllm/vllm-openai:v0.29.0"
},
"ovis-ocr2.py": {
"model_id": "ATH-MaaS/OvisOCR2",
"params": "0.9B",
"backend": "vllm",
"task": "ocr",
"output": "markdown + LaTeX + HTML tables",
"license": "apache-2.0",
"languages": {
"evidence": "not-stated"
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.29.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
},
"image": "vllm/vllm-openai:v0.29.0"
},
"ovis-ocr2-server.py": {
"model_id": "ATH-MaaS/OvisOCR2",
"params": "0.9B",
"backend": "vllm-server (in-job vllm serve + concurrent driver)",
"image": "vllm/vllm-openai:v0.22.1",
"task": "ocr",
"output": "markdown + LaTeX + HTML tables",
"license": "apache-2.0",
"notes": "Server-mode sibling of ovis-ocr2.py: ~1.7x its inference throughput, per-request failure isolation. See SERVING.md.",
"languages": {
"evidence": "not-stated"
},
"support": "legacy",
"support_note": "Use ovis-ocr2-saturate.py (same model and server, adaptive concurrency, resumable)."
},
"ovis-ocr2-saturate.py": {
"model_id": "ATH-MaaS/OvisOCR2",
"params": "0.9B",
"backend": "saturate (in-job vllm serve + saturate pump)",
"image": "vllm/vllm-openai:v0.22.1",
"task": "ocr",
"output": "markdown + LaTeX + HTML tables",
"license": "apache-2.0",
"notes": "saturate companion of ovis-ocr2-server.py: same model/prompt/sampling/post-processing; the driver half (adaptive concurrency, retries, resumable parquet output, durable error rows) is the saturate library. SERVING dict at the top of the script is the machine-readable tuning prior.",
"languages": {
"evidence": "not-stated"
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.22.1",
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
}
},
"xiaomi-ocr-0-saturate.py": {
"model_id": "xiaomi-research/Xiaomi-OCR-0",
"params": "0.8B",
"backend": "saturate (in-job vllm serve or SGLang + saturate pump)",
"image": "vllm/vllm-openai:v0.29.0",
"task": "ocr",
"output": "markdown + LaTeX + HTML tables",
"license": "apache-2.0",
"notes": "Single-stage whole-page mode. Card scores (96.83 OmniDocBench v1.6, 95.24 Real5, 87.94 Wild) come from Xiaomi's two-stage pipeline, which this recipe does not run. --engine sglang needs --image lmsysorg/sglang:v0.5.20-cu130. Measured on UFO scans: a10g-small 3.4 pages/s, rtx-pro-6000 9.7 pages/s, both about $0.08 per 1,000 pages excluding boot.",
"languages": {
"evidence": "not-stated"
},
"support": "core",
"support_note": "Works well on clean print. On small-print historical newspapers it can loop or insert Chinese characters, with finish_reason still stop; check tables for lost items and separators.",
"jobs": {
"image": "vllm/vllm-openai:v0.29.0",
"flavor": "a10g-small",
"timeout": "4h",
"secrets": [
"HF_TOKEN"
]
}
},
"lighton-ocr.py": {
"model_id": "lightonai/LightOnOCR-1B-1025",
"params": "1B",
"backend": "vllm",
"task": "ocr",
"output": "markdown",
"languages": {
"evidence": "named-list",
"count": 9,
"named": [
"en",
"fr",
"de",
"es",
"it",
"nl",
"pt",
"sv",
"da"
],
"notes": "Explicitly European / Latin-alphabet focused."
},
"support": "legacy",
"support_note": "LightOnOCR v1; use lighton-ocr2.py."
},
"lighton-ocr2.py": {
"model_id": "lightonai/LightOnOCR-2-1B",
"params": "1B",
"backend": "vllm",
"task": "ocr",
"output": "markdown",
"languages": {
"evidence": "named-list",
"count": 11,
"named": [
"en",
"fr",
"de",
"es",
"it",
"nl",
"pt",
"sv",
"da",
"zh",
"ja"
],
"notes": "Adds zh/ja over v1; declared, not benchmarked per language."
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.22.1",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
},
"image": "vllm/vllm-openai:v0.22.1"
},
"lighton-ocr3.py": {
"model_id": "lightonai/LightOnOCR-3-4B",
"params": "4B (or 0.8B via --model)",
"backend": "vllm",
"task": "ocr",
"output": "markdown; --mode grounding adds labelled layout blocks with 0-1000 boxes (layout_blocks JSON column)",
"license": "apache-2.0",
"languages": {
"evidence": "not-stated",
"notes": "LightOnOCR-3 cards give no language list (LightOnOCR-2 named 11)."
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.30.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
},
"image": "vllm/vllm-openai:v0.30.0",
"notes": "Also runs lightonai/LightOnOCR-3-0.8B. Not LightOnOCR-3-1B (Pixtral arch, different pins). vLLM 0.30.0 pinned: upstream reports 0.27.1 loops."
},
"lighton-ocr3-bucket.py": {
"model_id": "lightonai/LightOnOCR-3-4B",
"params": "4B (or 0.8B via --model)",
"backend": "vllm",
"task": "ocr",
"output": "one .md per image/PDF page in an output dir; --mode grounding adds a .json sidecar of layout blocks",
"license": "apache-2.0",
"languages": {
"evidence": "not-stated",
"notes": "LightOnOCR-3 cards give no language list (LightOnOCR-2 named 11)."
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.30.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
},
"image": "vllm/vllm-openai:v0.30.0",
"notes": "Bucket sibling of lighton-ocr3.py (same model, prompt, sampling). Input/output: hf://buckets/... URLs via bucketbag v0.3.1 (no mount) or directories. Reads .jp2; resume-by-skip."
},
"lighton-ocr2-server.py": {
"model_id": "lightonai/LightOnOCR-2-1B",
"params": "1B",
"backend": "vllm-server (in-job vllm serve + concurrent driver)",
"image": "vllm/vllm-openai:v0.22.1",
"task": "ocr",
"output": "markdown",
"notes": "Server-mode sibling of lighton-ocr2.py (the model card's own documented path): ~1.8x its inference throughput. See SERVING.md.",
"languages": {
"evidence": "named-list",
"count": 11,
"named": [
"en",
"fr",
"de",
"es",
"it",
"nl",
"pt",
"sv",
"da",
"zh",
"ja"
],
"notes": "Adds zh/ja over v1; declared, not benchmarked per language."
},
"support": "legacy",
"support_note": "Use lighton-ocr2-saturate.py (same model and server, adaptive concurrency, resumable)."
},
"lighton-ocr2-saturate.py": {
"model_id": "lightonai/LightOnOCR-2-1B",
"params": "1B",
"backend": "saturate (in-job vllm serve + saturate pump)",
"image": "vllm/vllm-openai:v0.22.1",
"task": "ocr",
"output": "markdown",
"notes": "saturate companion of lighton-ocr2-server.py: same model/message shape/sampling; the driver half (adaptive concurrency, retries, resumable parquet output, durable error rows) is the saturate library. SERVING dict at the top of the script is the machine-readable tuning prior.",
"languages": {
"evidence": "named-list",
"count": 11,
"named": [
"en",
"fr",
"de",
"es",
"it",
"nl",
"pt",
"sv",
"da",
"zh",
"ja"
],
"notes": "Adds zh/ja over v1; declared, not benchmarked per language."
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.22.1",
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
}
},
"hunyuan-ocr.py": {
"model_id": "tencent/HunyuanOCR",
"revision": "f6af82ee007fe6091b29fb3bb287b491ead41c82",
"params": "1B",
"backend": "vllm",
"task": "ocr",
"output": "markdown",
"license": "Hunyuan Community License (excludes EU/UK/KR)",
"languages": {
"evidence": "multilingual-unspecified"
},
"support": "unsupported",
"support_note": "Broken (2026-09-23): the pinned 1.0 revision fails on current vLLM/transformers. Use hunyuan-ocr-1.5.py."
},
"hunyuan-ocr-1.5.py": {
"model_id": "tencent/HunyuanOCR",
"params": "1B",
"backend": "vllm",
"task": "ocr | 12 task types",
"output": "markdown",
"license": "Hunyuan Community License (excludes EU/UK/KR)",
"languages": {
"evidence": "multilingual-unspecified",
"notes": "Explicitly targets low-resource + ancient-script OCR as a design goal; nothing enumerated."
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.24.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
},
"image": "vllm/vllm-openai:v0.24.0"
},
"dots-ocr.py": {
"model_id": "rednote-hilab/dots.ocr",
"params": "1.7B",
"backend": "vllm",
"task": "ocr + layout",
"output": "markdown",
"languages": {
"evidence": "count-claim",
"count": 100,
"notes": "Explicit low-resource claim backed by in-house dots.ocr-bench (1,493 PDFs across 100 languages); aggregate scores only."
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.29.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
},
"image": "vllm/vllm-openai:v0.29.0"
},
"firered-ocr.py": {
"model_id": "FireRedTeam/FireRed-OCR",
"params": "2.1B",
"backend": "vllm",
"task": "ocr",
"output": "markdown",
"license": "apache-2.0",
"languages": {
"evidence": "not-stated"
},
"support": "legacy",
"support_note": "Works; low adoption."
},
"abot-ocr.py": {
"model_id": "acvlab/ABot-OCR",
"params": "2B",
"backend": "vllm",
"image": "vllm/vllm-openai",
"task": "ocr",
"output": "markdown (text, LaTeX, HTML tables)",
"languages": {
"evidence": "not-stated"
},
"support": "legacy",
"support_note": "Works; low adoption and no declared model licence."
},
"nanonets-ocr.py": {
"model_id": "nanonets/Nanonets-OCR-s",
"params": "2B",
"backend": "vllm",
"task": "ocr",
"output": "markdown (LaTeX, tables, forms)",
"languages": {
"evidence": "english-only"
},
"support": "legacy",
"support_note": "Nanonets-OCR-s; use nanonets-ocr2.py."
},
"dots-mocr.py": {
"model_id": "rednote-hilab/dots.mocr",
"params": "3B",
"backend": "vllm",
"task": "ocr | 8 prompt modes incl. SVG, layout + bbox",
"output": "markdown",
"languages": {
"evidence": "multilingual-unspecified"
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.29.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
},
"image": "vllm/vllm-openai:v0.29.0"
},
"nanonets-ocr2.py": {
"model_id": "nanonets/Nanonets-OCR2-3B",
"params": "3B",
"backend": "vllm",
"image": "vllm/vllm-openai:v0.10.2",
"task": "ocr",
"output": "markdown",
"languages": {
"evidence": "named-list",
"count": "11 named + 'many more'",
"named": [
"en",
"zh",
"fr",
"es",
"pt",
"de",
"it",
"ru",
"ja",
"ko",
"ar"
],
"notes": "Only card claiming multilingual handwriting; list illustrative, no per-language benchmark."
},
"support": "legacy",
"support_note": "Works on its pinned vllm/vllm-openai:v0.10.2 image; not header-migrated."
},
"deepseek-ocr-vllm.py": {
"model_id": "deepseek-ai/DeepSeek-OCR",
"params": "4B",
"backend": "vllm",
"task": "ocr | 5 resolution + 5 prompt modes",
"output": "markdown",
"license": "mit",
"languages": {
"evidence": "multilingual-unspecified"
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.29.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
},
"image": "vllm/vllm-openai:v0.29.0"
},
"jina-ocr-v1.py": {
"model_id": "jinaai/jina-ocr-v1",
"model_name": "jina-ocr-v1",
"params": "3.4B (MoE, ~570M active)",
"backend": "vllm",
"task": "ocr",
"output": "markdown",
"license": "cc-by-nc-4.0",
"notes": "DeepSeek-OCR fine-tune with a FastMTP speculative-decoding head; offline vLLM only (the checkpoint's register() must run in-process, no vllm serve path yet). Card: 91.14 OmniDocBench v1.6, 83.4 olmOCR-Bench.",
"languages": {
"evidence": "multilingual-unspecified",
"notes": "Card metadata says multilingual; no per-language list or evidence in the card body."
},
"support": "unsupported",
"support_note": "Broken (2026-09-23): vLLM 0.30 DeepEncoder kernel error; pinning vllm/vllm-openai:v0.29.0 likely fixes it (as for deepseek-ocr-vllm.py). CC-BY-NC-4.0 model.",
"jobs": {
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
],
"env": {
"UV_TORCH_BACKEND": "auto"
}
}
},
"deepseek-ocr.py": {
"model_id": "deepseek-ai/DeepSeek-OCR",
"params": "4B",
"backend": "transformers",
"task": "ocr",
"output": "markdown",
"license": "mit",
"languages": {
"evidence": "multilingual-unspecified"
},
"support": "unsupported",
"support_note": "Broken (2026-09-23): every output row is the string 'None' (model.infer returns None). Use deepseek-ocr-vllm.py."
},
"deepseek-ocr2-vllm.py": {
"model_id": "deepseek-ai/DeepSeek-OCR-2",
"params": "3B",
"backend": "vllm (nightly)",
"image": "vllm/vllm-openai:v0.29.0",
"task": "ocr",
"output": "markdown",
"license": "apache-2.0",
"languages": {
"evidence": "multilingual-unspecified"
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.29.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
}
},
"unlimited-ocr2.py": {
"model_id": "baidu/Unlimited-OCR-2",
"params": "3.3B (MoE, 0.5B active)",
"backend": "saturate (in-job SGLang from Baidu's wheel + saturate pump)",
"image": "lmsysorg/sglang:v0.5.20-cu130",
"task": "ocr (multi-page documents: folders of page images; cross-page continuations)",
"output": "markdown per page + layout-grounded raw, one row per 16-page window",
"license": "mit",
"notes": "Sends 16 consecutive pages per request so the model keeps context across page breaks. Needs a >= 48 GB GPU (24 GB cards thrash the KV cache). Measured on 1771 Britannica: rtx-pro-6000 ~1,570 tok/s end to end. About 2% of dense 18th-century pages loop; unlimited-ocr2-saturate.py repairs them.",
"languages": {
"evidence": "multilingual-unspecified"
},
"support": "core",
"jobs": {
"image": "lmsysorg/sglang:v0.5.20-cu130",
"flavor": "rtx-pro-6000",
"timeout": "4h",
"secrets": [
"HF_TOKEN"
]
}
},
"unlimited-ocr2-saturate.py": {
"model_id": "baidu/Unlimited-OCR-2",
"params": "3.3B (MoE, 0.5B active)",
"backend": "saturate (in-job SGLang from Baidu's wheel + saturate pump)",
"image": "lmsysorg/sglang:v0.5.20-cu130",
"task": "ocr (multi-page documents: book volumes, PDFs; cross-page continuations)",
"output": "markdown per page and per window + layout-grounded raw",
"license": "mit",
"notes": "Tuned version of unlimited-ocr2.py: repair pass for looping pages (fixed 28/33 in testing), PDF input, Baidu's own Markdown post-processing, speed flags.",
"languages": {
"evidence": "multilingual-unspecified"
},
"support": "experimental",
"support_note": "Use unlimited-ocr2.py unless you need the repair pass, PDFs or the extras.",
"jobs": {
"image": "lmsysorg/sglang:v0.5.20-cu130",
"flavor": "rtx-pro-6000",
"timeout": "4h",
"secrets": [
"HF_TOKEN"
]
}
},
"unlimited-ocr2-live.py": {
"model_id": "baidu/Unlimited-OCR-2",
"params": "3.3B (MoE, 0.5B active)",
"backend": "saturate (in-job SGLang from Baidu's wheel + saturate pump)",
"image": "lmsysorg/sglang:v0.5.20-cu130",
"task": "ocr (multi-page documents) with a live view of the run",
"output": "markdown per page + layout-grounded raw, one row per 16-page window",
"license": "mit",
"notes": "unlimited-ocr2.py plus --live-port: a dashboard served from the Job (start-up steps, the page being read with its text streaming in, finished pages, pages/min, GPU load) for watching or screen-recording a run. Measured on 1,456 Wellcome Library pages: ~130 pages/min on rtx-pro-6000.",
"languages": {
"evidence": "multilingual-unspecified"
},
"support": "experimental",
"support_note": "Use unlimited-ocr2.py for normal runs; this adds a live view for demos.",
"jobs": {
"image": "lmsysorg/sglang:v0.5.20-cu130",
"flavor": "rtx-pro-6000",
"timeout": "4h",
"secrets": [
"HF_TOKEN"
]
}
},
"minicpm-ocr.py": {
"model_id": "openbmb/MiniCPM-OCR-1.0",
"params": "1.3B",
"backend": "in-job vLLM (nightly) + OpenBMB's minicpm-ocr SDK (PP-DocLayoutV3 layout, per-region prompts, Markdown assembly)",
"image": "cr.hf.co/davanstrien/minicpm-ocr:1.0-vllm-5d35ce3",
"task": "ocr (page images -> Markdown; tables as HTML, formulas as LaTeX); Hub dataset or Bucket/folder I/O",
"output": "dataset mode: markdown + markdown_status columns; Bucket/folder mode: one .md per image (failed pages not written, re-run retries)",
"license": "apache-2.0",
"notes": "A region reader behind a layout model, run through the vendor SDK unchanged. Custom public image (Dockerfile in images/minicpm-ocr/): the architecture is in vLLM main but not a release, and the SDK needs PaddlePaddle in its own venv. DSpark speculative decoding on by default (--no-dspark). Bucket I/O via bucketbag v0.3.1 (no mount), resume-by-skip. Smoke-tested 2026-10-27 on BHL pages (dataset mode) and 1771 Britannica .jp2 (Bucket mode), a10g-large.",
"languages": {
"evidence": "named-list",
"count": 2,
"named": [
"zh",
"en"
],
"notes": "Card: 'Primary languages: Chinese and English'; card metadata lists en, zh. Card examples include historical scans; fine-tuning guide uses French handwriting."
},
"support": "experimental",
"jobs": {
"image": "cr.hf.co/davanstrien/minicpm-ocr:1.0-vllm-5d35ce3",
"flavor": "a10g-large",
"timeout": "2h",
"secrets": [
"HF_TOKEN"
]
}
},
"unlimited-ocr-vllm.py": {
"model_id": "baidu/Unlimited-OCR",
"params": "3.3B",
"backend": "vllm",
"image": "vllm/vllm-openai:unlimited-ocr",
"task": "ocr (layout-grounded markdown; single-image batch)",
"output": "markdown",
"license": "mit",
"languages": {
"evidence": "multilingual-unspecified"
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:unlimited-ocr",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
}
},
"nuextract3.py": {
"model_id": "numind/NuExtract3",
"params": "4B",
"backend": "vllm",
"image": "vllm/vllm-openai:v0.29.0",
"task": "ocr | schema-guided extraction",
"output": "markdown or JSON",
"languages": {
"evidence": "multilingual-unspecified"
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.29.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
}
},
"qianfan-ocr.py": {
"model_id": "baidu/Qianfan-OCR",
"params": "4.7B",
"backend": "vllm",
"task": "ocr",
"output": "markdown",
"languages": {
"evidence": "count-claim",
"count": 192,
"notes": "Headline claim; no list or per-language numbers."
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.29.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
},
"image": "vllm/vllm-openai:v0.29.0"
},
"olmocr2-vllm.py": {
"model_id": "allenai/olmOCR-2-7B-1025-FP8",
"params": "7B",
"backend": "vllm",
"task": "ocr",
"output": "markdown",
"languages": {
"evidence": "english-only"
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.29.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
},
"image": "vllm/vllm-openai:v0.29.0"
},
"rolm-ocr.py": {
"model_id": "reducto/RolmOCR",
"params": "7B",
"backend": "vllm",
"task": "ocr",
"output": "plain text",
"languages": {
"evidence": "not-stated"
},
"support": "unsupported",
"support_note": "Broken (2026-09-23): no KV cache room on a 24 GB L4 with the default settings. Use olmocr2-vllm.py."
},
"numarkdown-ocr.py": {
"model_id": "numind/NuMarkdown-8B-Thinking",
"params": "8B",
"backend": "vllm",
"task": "ocr (reasoning)",
"output": "markdown",
"languages": {
"evidence": "not-stated"
},
"support": "legacy",
"support_note": "8B reasoning model; works, little used."
},
"lfm2-vl-extract.py": {
"model_id": "LiquidAI/LFM2.5-VL-1.6B-Extract",
"params": "1.6B",
"backend": "vllm",
"image": "vllm/vllm-openai",
"task": "schema-guided extraction (image -> JSON)",
"output": "JSON",
"languages": {
"evidence": "english-only",
"notes": "Vision card declares en only; the text sibling's 9-language list does NOT carry over."
},
"support": "legacy",
"support_note": "Works; nuextract3.py and lift-extract.py cover image-to-JSON extraction."
},
"lfm2-extract.py": {
"model_id": "LiquidAI/LFM2-1.2B-Extract",
"params": "1.2B",
"backend": "vllm",
"image": "vllm/vllm-openai:v0.29.0",
"task": "schema-guided extraction (text -> structured); chain after any OCR recipe",
"output": "JSON / XML / YAML",
"languages": {
"evidence": "named-list",
"count": 9,
"named": [
"en",
"ar",
"zh",
"fr",
"de",
"ja",
"ko",
"pt",
"es"
],
"notes": "Text-only model (not OCR). Card prose names 9 incl. Portuguese; card YAML lists 8 (pt missing)."
},
"support": "tool",
"jobs": {
"image": "vllm/vllm-openai:v0.29.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
}
},
"lift-extract.py": {
"model_id": "datalab-to/lift",
"params": "9B",
"backend": "transformers | vllm",
"task": "schema-guided extraction (image or multi-page PDF -> JSON)",
"output": "JSON",
"license": "modified OpenRAIL-M",
"languages": {
"evidence": "not-stated"
},
"support": "core",
"jobs": {
"flavor": "a100-large",
"secrets": [
"HF_TOKEN"
]
}
},
"pp-doclayout.py": {
"model_id": "PaddlePaddle/PP-DocLayout-L",
"params": "123M",
"backend": "paddleocr",
"task": "layout detection (no text extraction)",
"output": "bounding boxes + region classes",
"languages": {
"evidence": "not-applicable"
},
"support": "tool",
"jobs": {
"flavor": "t4-small",
"secrets": [
"HF_TOKEN"
]
}
},
"glm-ocr-v2.py": {
"variant_of": "glm-ocr.py",
"notes": "Adds checkpoint/resume for very large jobs.",
"support": "legacy",
"support_note": "Superseded by glm-ocr.py for most runs; kept for its resume support."
},
"glm-ocr-bucket.py": {
"variant_of": "glm-ocr.py",
"notes": "Reads images/PDFs from a mounted bucket, writes one .md per page.",
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.29.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
},
"image": "vllm/vllm-openai:v0.29.0"
},
"falcon-ocr-bucket.py": {
"variant_of": "falcon-ocr.py",
"notes": "Reads images/PDFs from a mounted bucket, writes one .md per page.",
"support": "legacy",
"support_note": "Falcon-OCR v1; v1.5 work is in PR #110. --max-samples counts files, not pages."
},
"surya-ocr-bucket.py": {
"variant_of": "surya-ocr.py",
"notes": "Bucket-to-bucket structured OCR (--io-mode mount|copy), resumable.",
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.20.1",
"python": "/usr/local/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/site-packages"
},
"flavor": "a10g-small",
"secrets": [
"HF_TOKEN"
]
},
"image": "vllm/vllm-openai:v0.20.1"
},
"ocr-vllm-judge.py": {
"model_id": "configurable",
"task": "pairwise OCR-quality judging (VLM-as-judge)",
"languages": {
"evidence": "not-applicable"
},
"support": "tool",
"jobs": {
"image": "vllm/vllm-openai:v0.29.0",
"python": "/usr/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/dist-packages"
},
"flavor": "a100-large",
"secrets": [
"HF_TOKEN"
]
},
"image": "vllm/vllm-openai:v0.29.0"
}
}
{
"_meta": {
"description": "Machine-readable catalog of the OCR recipes in this directory: script -> model, params, backend, support level, the tested Jobs launch config (the script's [tool.hf-jobs] header), and what the model's own card claims about language coverage. Human-readable language notes: LANGUAGES.md. Language evidence levels, strongest first: per-language-benchmark | count-claim | named-list | multilingual-unspecified | english-only | not-stated | not-applicable. Claims come from model cards (compiled 2026-07-15) - treat as claims, not guarantees.",
"updated": "2026-10-30",
"run_pattern": "hf jobs uv run https://huggingface.co/datasets/uv-scripts/ocr/raw/main/<script> <input-dataset> <output-dataset> (hf CLI >= 1.32 reads image/flavor/secrets from the script's [tool.hf-jobs] header; add --timeout for long runs. Scripts without a `jobs` entry need --flavor/-s HF_TOKEN (and any --image in their docstring).)",
"support_levels": {
"core": "Maintained. Smoke-tested flagless on HF Jobs on 2026-09-23/24 with the pinned launch config in `jobs`.",
"tool": "Maintained companion (judge, layout, text extraction), not a page-OCR recipe.",
"legacy": "Works (smoke-tested 2026-09-23) but superseded or little used; no header yet. See support_note for the alternative.",
"unsupported": "Known broken as of 2026-09-23; kept in the repo for now. See support_note.",
"experimental": "Works in testing on HF Jobs with the launch config in `jobs`, but changes often and may break. See support_note for the stable alternative."
}
},
"tesseract-ocr.py": {
"model_id": null,
"model_name": "Tesseract 5",
"params": null,
"backend": "pytesseract (CPU)",
"task": "ocr",
"output": "plain text",
"license": "apache-2.0",
"languages": {
"evidence": "named-list",
"count": "125 traineddata packs (~100+ languages + script models)",
"list_url": "https://github.com/tesseract-ocr/tessdata_best",
"notes": "Broadest named coverage; script models (Latin, Cyrillic, Devanagari, ...) cover languages without a dedicated pack. Legacy baseline quality."
},
"support": "core",
"jobs": {
"flavor": "cpu-upgrade",
"secrets": [
"HF_TOKEN"
]
}
},
"pp-ocrv6.py": {
"model_id": "PaddlePaddle/PP-OCRv6",
"params": "1.5M-34.5M",
"backend": "paddleocr",
"task": "ocr",
"output": "plain text",
"license": "apache-2.0",
"languages": {
"evidence": "count-claim",
"count": 48
},
"support": "core",
"jobs": {
"flavor": "t4-small",
"secrets": [
"HF_TOKEN"
]
}
},
"falcon-ocr.py": {
"model_id": "tiiuae/Falcon-OCR",
"params": "0.3B",
"backend": "falcon-perception",
"task": "ocr",
"output": "markdown",
"license": "apache-2.0",
"languages": {
"evidence": "not-stated"
},
"support": "legacy",
"support_note": "Falcon-OCR v1; v1.5 work is in PR #110."
},
"smoldocling-ocr.py": {
"model_id": "ds4sd/SmolDocling-256M-preview",
"params": "256M",
"backend": "transformers",
"task": "ocr",
"output": "DocTags",
"languages": {
"evidence": "english-only"
},
"support": "unsupported",
"support_note": "Broken (2026-09-23): the DocTags-to-markdown converter is a stub, so rows are raw DocTags."
},
"surya-ocr.py": {
"model_id": "datalab-to/surya-ocr-2",
"params": "0.65B",
"backend": "vllm",
"image": "vllm/vllm-openai:v0.20.1",
"task": "ocr | layout | table",
"output": "markdown + surya_blocks (per-block HTML, bboxes, reading order)",
"license": "modified OpenRAIL-M",
"languages": {
"evidence": "per-language-benchmark",
"count": 91,
"benchmark_url": "https://github.com/datalab-to/surya/blob/master/static/docs/multilingual.md",
"notes": "38/91 languages score >=90%, 76/91 >=80%; weakest of top-15: Arabic 72.7%, Vietnamese 73.2%."
},
"support": "core",
"jobs": {
"image": "vllm/vllm-openai:v0.20.1",
"python": "/usr/local/bin/python3",
"env": {
"PYTHONPATH": "/usr/local/lib/python3.12/site-packages"
},