model_registry.json

{
  "models": [
    {
      "id": "qwen3_vl_8b_instruct",
      "name": "Qwen3-VL-8B-Instruct",
      "model_id": "Qwen/Qwen3-VL-8B-Instruct",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "enabled": true,
      "prompt": "Transcribe every piece of printed and handwritten text from this document image. Keep paragraph/line layout when obvious. Return plain text only.",
      "parameters": {
        "max_new_tokens": 1024,
        "temperature": 0.1,
        "top_p": 0.95,
        "do_sample": false
      },
      "notes": "Best default config for clean extraction and low hallucination."
    },
    {
      "id": "qwen3_vl_4b_instruct",
      "name": "Qwen3-VL-4B-Instruct",
      "model_id": "Qwen/Qwen3-VL-4B-Instruct",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "enabled": true,
      "prompt": "Transcribe every piece of printed and handwritten text from this Hebrew document image. Keep line breaks and text order.",
      "parameters": {
        "max_new_tokens": 1024,
        "temperature": 0.1,
        "top_p": 0.95,
        "do_sample": false
      },
      "notes": "Smallest Qwen3-VL baseline for fast OCR baselining."
    },
    {
      "id": "qwen3_vl_8b_thinking",
      "name": "Qwen3-VL-8B-Thinking",
      "model_id": "Qwen/Qwen3-VL-8B-Thinking",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "enabled": true,
      "prompt": "Use deliberate reasoning to first recover layout and then output final OCR text exactly as visible. Hide reasoning and return only plain text.",
      "parameters": {
        "max_new_tokens": 1536,
        "temperature": 0.0,
        "top_p": 1.0,
        "extra_body": {
          "enable_thinking": true
        }
      },
      "notes": "Thinking mode usually helps on challenging handwriting and dense layouts."
    },
    {
      "id": "qwen3_vl_4b_thinking",
      "name": "Qwen3-VL-4B-Thinking",
      "model_id": "Qwen/Qwen3-VL-4B-Thinking",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "enabled": true,
      "prompt": "Use a deliberate OCR pass for printed and handwritten Hebrew text, then output only the final transcription.",
      "parameters": {
        "max_new_tokens": 1400,
        "temperature": 0.0,
        "top_p": 1.0,
        "extra_body": {
          "enable_thinking": true
        }
      },
      "notes": "4B thinking pass for hard handwriting/low-contrast lines."
    },
    {
      "id": "qwen3_vl_30b_a3b_instruct",
      "name": "Qwen3-VL-30B-A3B-Instruct",
      "model_id": "Qwen/Qwen3-VL-30B-A3B-Instruct",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "enabled": true,
      "prompt": "OCR this Hebrew document image and return plain text only, preserving order, lines and spacing.",
      "parameters": {
        "max_new_tokens": 2048,
        "temperature": 0.0,
        "top_p": 0.95
      },
      "notes": "Higher-capacity Qwen3-VL checkpoint for benchmark ceiling."
    },
    {
      "id": "qwen3_vl_30b_a3b_thinking",
      "name": "Qwen3-VL-30B-A3B-Thinking",
      "model_id": "Qwen/Qwen3-VL-30B-A3B-Thinking",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "enabled": true,
      "prompt": "Think through layout and character ambiguity first, then output final OCR text only.",
      "parameters": {
        "max_new_tokens": 3072,
        "temperature": 0.0,
        "top_p": 1.0,
        "extra_body": {
          "enable_thinking": true
        }
      },
      "notes": "Thinking-oriented 30B setup for the hardest mixed handwriting cases."
    },
    {
      "id": "nemotron_ocr_v2",
      "name": "Nemotron-OCR-v2",
      "model_id": "nvidia/nemotron-ocr-v2",
      "provider": "local_transformer",
      "enabled": true,
      "prompt": "Transcribe all printed and handwritten text in this Hebrew document image. Return only plain text.",
      "parameters": {
        "max_new_tokens": 1536,
        "temperature": 0.0,
        "top_p": 0.95
      },
      "notes": "Multilingual OCR model with strong layout and text recognition behavior."
    },
    {
      "id": "gemma4_e4b_instruct",
      "name": "Gemma4 0.4B (E4B) Instruct",
      "model_id": "google/gemma-4-E4B-it",
      "provider": "local_transformer",
      "enabled": true,
      "prompt": "You are doing strict OCR. Extract printed and handwritten Hebrew/English text only. Keep every line and punctuation as shown.",
      "parameters": {
        "max_new_tokens": 1024,
        "temperature": 0.0,
        "top_p": 0.9
      },
      "notes": "Gemma4 entry for the 4B family."
    },
    {
      "id": "gemma4_e4b_thinking",
      "name": "Gemma4 0.4B (E4B) Thinking",
      "model_id": "google/gemma-4-E4B-it",
      "provider": "local_transformer",
      "enabled": true,
      "prompt": "Think carefully through layout, then output only the final document transcription with line breaks preserved.",
      "parameters": {
        "max_new_tokens": 1536,
        "temperature": 0.0,
        "top_p": 1.0,
        "extra_body": {
          "enable_thinking": true
        }
      },
      "notes": "Same checkpoint with higher budget and thinking-oriented system prompt."
    },
    {
      "id": "gemma4_12b_instruct",
      "name": "Gemma4 12B Instruct",
      "model_id": "google/gemma-4-12B-it",
      "provider": "local_transformer",
      "enabled": true,
      "prompt": "OCR only. Return all readable printed and handwritten text in reading order. Do not describe anything else.",
      "parameters": {
        "max_new_tokens": 1536,
        "temperature": 0.0,
        "top_p": 0.95
      },
      "notes": "Best default for instruction mode."
    },
    {
      "id": "gemma4_12b_thinking",
      "name": "Gemma4 12B Thinking",
      "model_id": "google/gemma-4-12B-it",
      "provider": "local_transformer",
      "enabled": true,
      "prompt": "Use internal planning for document structure, then output the final OCR text only.",
      "parameters": {
        "max_new_tokens": 2048,
        "temperature": 0.0,
        "top_p": 1.0,
        "extra_body": {
          "enable_thinking": true
        }
      },
      "notes": "Thinking-oriented OCR prompt with larger decoding budget."
    },
    {
      "id": "gemma4_26b4b_instruct",
      "name": "Gemma4 26B-A4B Instruct",
      "model_id": "google/gemma-4-26B-A4B-it",
      "provider": "local_transformer",
      "enabled": true,
      "prompt": "Transcribe the document exactly. Keep whitespace between words/lines and don't clean the content.",
      "parameters": {
        "max_new_tokens": 2048,
        "temperature": 0.0,
        "top_p": 0.9
      },
      "notes": "Larger Gemma4-A4B OCR baseline."
    },
    {
      "id": "gemma4_26b4b_thinking",
      "name": "Gemma4 26B-A4B Thinking",
      "model_id": "google/gemma-4-26B-A4B-it",
      "provider": "local_transformer",
      "enabled": true,
      "prompt": "Resolve handwriting and layout ambiguities first, then return only the final cleaned OCR output.",
      "parameters": {
        "max_new_tokens": 2560,
        "temperature": 0.0,
        "top_p": 1.0,
        "extra_body": {
          "enable_thinking": true
        }
      },
      "notes": "Thinking prompt + higher budget."
    },
    {
      "id": "gemma4_31b_instruct",
      "name": "Gemma4 31B Instruct",
      "model_id": "google/gemma-4-31B-it",
      "provider": "local_transformer",
      "enabled": true,
      "prompt": "Strict OCR pass: output only text from the image, maintaining line structure.",
      "parameters": {
        "max_new_tokens": 2048,
        "temperature": 0.0,
        "top_p": 0.95
      },
      "notes": "Highest-capacity Gemma4 variant in this panel."
    },
    {
      "id": "gemma4_31b_thinking",
      "name": "Gemma4 31B Thinking",
      "model_id": "google/gemma-4-31B-it",
      "provider": "local_transformer",
      "enabled": true,
      "prompt": "Do internal chain-of-thought for OCR confidence, then output only the final transcription.",
      "parameters": {
        "max_new_tokens": 3072,
        "temperature": 0.0,
        "top_p": 1.0,
        "extra_body": {
          "enable_thinking": true
        }
      },
      "notes": "Thinking-oriented prompt for hard mixed-direction text."
    },
    {
      "id": "chandra2",
      "name": "Chandra2 (datalab-to/chandra-ocr-2)",
      "model_id": "datalab-to/chandra-ocr-2",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "Output only the exact text from the uploaded document, including handwritten and printed parts.",
      "parameters": {
        "max_new_tokens": 1024,
        "temperature": 0.0
      },
      "notes": "Document OCR and layout-focused model."
    },
    {
      "id": "ronylicha_gigapdf",
      "name": "ronylicha/gigapdf-ocr-hebrew",
      "model_id": "ronylicha/gigapdf-ocr-hebrew",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "Read the Hebrew document image and return only transcribed text.",
      "parameters": {
        "max_new_tokens": 1024,
        "temperature": 0.0
      },
      "notes": "Hebrew-first OCR model; sometimes better with Hebrew-only content."
    },
    {
      "id": "qunie_v7_mini",
      "name": "Qunie-V7-mini",
      "model_id": "liskcell/Qunie-V7-mini",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "Strict OCR extraction. Output line-by-line text from the whole image.",
      "parameters": {
        "max_new_tokens": 1536,
        "temperature": 0.1,
        "top_p": 0.96
      },
      "notes": "Balanced performance/speed config."
    },
    {
      "id": "openmlkitocr",
      "name": "OpenMLKitOCR",
      "model_id": "0cve0/OpenMLKitOCR",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "OCR the document and return readable text as-is.",
      "parameters": {
        "max_new_tokens": 1024,
        "temperature": 0.0
      },
      "notes": "Mobile-oriented OCR stack from ML Kit; tuned for mixed scans."
    },
    {
      "id": "tzefa_word_ocr_trocr",
      "name": "Tzefa-Word-OCR-TrOCR",
      "model_id": "WARAJA/Tzefa-Word-OCR-TrOCR",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "Do OCR on this image. Return only transcribed words in reading order.",
      "parameters": {
        "max_new_tokens": 1024,
        "temperature": 0.0
      },
      "notes": "Word-level model behavior for clean printed text lines."
    },
    {
      "id": "qunie_v7_pico",
      "name": "Qunie-V7-Pico",
      "model_id": "liskcell/Qunie-V7-Pico",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "Extract all readable text. Prioritize transcription correctness over grammar cleanup.",
      "parameters": {
        "max_new_tokens": 1024,
        "temperature": 0.15,
        "top_p": 0.97
      },
      "notes": "Compact OCR model for faster response."
    },
    {
      "id": "aya_vision_32b",
      "name": "Aya Vision 32B",
      "model_id": "CohereLabs/aya-vision-32b",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "enabled": true,
      "prompt": "Return plain OCR output only from the image, including punctuation and line breaks.",
      "parameters": {
        "max_new_tokens": 2048,
        "temperature": 0.0,
        "top_p": 0.95
      },
      "notes": "High-capacity vision model; strong on multilingual text."
    },
    {
      "id": "aya_vision_8b",
      "name": "Aya Vision 8B",
      "model_id": "CohereLabs/aya-vision-8b",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "enabled": true,
      "prompt": "Return only text from the document image; preserve paragraph breaks where clear.",
      "parameters": {
        "max_new_tokens": 1536,
        "temperature": 0.0,
        "top_p": 0.95
      },
      "notes": "Balanced multilingual OCR speed/quality."
    },
    {
      "id": "phi4_multimodal",
      "name": "Phi-4-Multimodal-Instruct",
      "model_id": "microsoft/Phi-4-multimodal-instruct",
      "provider": "local_transformer",
      "pipeline_task": "image-text-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "OCR-only transcription. Output exactly what is printed/handwritten in the image.",
      "parameters": {
        "max_new_tokens": 1536,
        "temperature": 0.0,
        "top_p": 0.95
      },
      "notes": "Keep short and factual output."
    },
    {
      "id": "surya2",
      "name": "surya2 (datalab-to/surya-ocr-2)",
      "model_id": "datalab-to/surya-ocr-2",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "Return only OCR output from this document image, preserving line breaks and order.",
      "parameters": {
        "max_new_tokens": 1536,
        "temperature": 0.0,
        "top_p": 0.95
      },
      "notes": "Specialized OCR pipeline with layout awareness."
    },
    {
      "id": "paddleocr_vl",
      "name": "PaddleOCR-VL",
      "model_id": "PaddlePaddle/PaddleOCR-VL",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "Transcribe all readable printed and handwritten text from this document image in order.",
      "parameters": {
        "max_new_tokens": 4096,
        "temperature": 0.0
      },
      "notes": "PaddleOCR-VL base checkpoint."
    },
    {
      "id": "paddleocr_vl_15",
      "name": "PaddleOCR-VL 1.5",
      "model_id": "PaddlePaddle/PaddleOCR-VL-1.5",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "Transcribe all readable printed and handwritten text from this document image in order.",
      "parameters": {
        "max_new_tokens": 4096,
        "temperature": 0.0
      },
      "notes": "PaddleOCR-VL family variant."
    },
    {
      "id": "paddleocr_vl_16",
      "name": "PaddleOCR-VL 1.6",
      "model_id": "PaddlePaddle/PaddleOCR-VL-1.6",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "Transcribe all readable printed and handwritten text from this document image in order.",
      "parameters": {
        "max_new_tokens": 4096,
        "temperature": 0.0
      },
      "notes": "PaddleOCR-VL family variant."
    },
    {
      "id": "deepseek_ocr_1",
      "name": "DeepSeek OCR-1",
      "model_id": "deepseek-ai/DeepSeek-OCR",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "Transcribe all printed and handwritten text from this Hebrew document image. Output only plain text.",
      "parameters": {
        "max_new_tokens": 4096,
        "temperature": 0.0
      },
      "notes": "DeepSeek OCR-1 family baseline."
    },
    {
      "id": "deepseek_ocr_2",
      "name": "DeepSeek OCR-2",
      "model_id": "deepseek-ai/DeepSeek-OCR-2",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "Transcribe all printed and handwritten text from this Hebrew document image. Preserve line breaks as visible.",
      "parameters": {
        "max_new_tokens": 4096,
        "temperature": 0.0
      },
      "notes": "DeepSeek OCR-2 family OCR variant."
    },
    {
      "id": "tesseract_latest",
      "name": "Tesseract (local latest)",
      "model_id": "tesseract",
      "provider": "tesseract",
      "enabled": true,
      "parameters": {
        "lang": "heb+eng",
        "oem": 1,
        "psm": 6,
        "extra_config": "--dpi 300 --psm 6"
      },
      "prompt": "N/A",
      "notes": "Always available local baseline for Hebrew/English and free."
    },
    {
      "id": "cyttic_exp10_trocr_hebrew_matan_full",
      "name": "cyttic/exp10-trocr-hebrew-matan-full",
      "model_id": "cyttic/exp10-trocr-hebrew-matan-full",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "Transcribe all readable printed and handwritten Hebrew text from this document image. Return plain text only, preserve line breaks and spacing.",
      "parameters": {
        "max_new_tokens": 4096,
        "temperature": 0.0,
        "top_p": 0.95
      },
      "notes": "Added for Hebrew-focused TrOCR-style local OCR benchmarking."
    },
    {
      "id": "cyttic_exp23_directfit_unfrozen",
      "name": "cyttic/exp23-directfit-unfrozen",
      "model_id": "cyttic/exp23-directfit-unfrozen",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "OCR this document image and extract all printed and handwritten Hebrew text faithfully. Return only the plain transcription.",
      "parameters": {
        "max_new_tokens": 4096,
        "temperature": 0.0,
        "top_p": 0.95
      },
      "notes": "Added for offline direct-fit multilingual document OCR benchmarking."
    },
    {
      "id": "cyttic_heb_verifier17_connected",
      "name": "cyttic/heb-verifier17-connected",
      "model_id": "cyttic/heb-verifier17-connected",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "Return only the transcribed text from the uploaded document image (Hebrew document with printed and handwritten text). Preserve line order.",
      "parameters": {
        "max_new_tokens": 4096,
        "temperature": 0.0,
        "top_p": 0.95
      },
      "notes": "Added for verifier-focused Hebrew OCR comparison run."
    },
    {
      "id": "cyttic_exp26_composed1m",
      "name": "cyttic/exp26-composed1m",
      "model_id": "cyttic/exp26-composed1m",
      "provider": "local_transformer",
      "pipeline_task": "image-to-text",
      "trust_remote_code": true,
      "enabled": true,
      "prompt": "Transcribe exactly what is visible in the image, including Hebrew text lines and mixed-direction fragments.",
      "parameters": {
        "max_new_tokens": 4096,
        "temperature": 0.0,
        "top_p": 0.95
      },
      "notes": "Added for composed document OCR benchmark."
    }
  ]
}