CLAUDE.md

OCR Scripts - Development Notes

Active Scripts

DeepSeek-OCR v1 (deepseek-ocr-vllm.py)

✅ Production Ready

• Fully supported by vLLM
• Fast batch processing
• Tested and working on HF Jobs

LightOnOCR-2-1B (lighton-ocr2.py)

✅ Production Ready (Fixed 2026-01-29)

Status: Working with vLLM nightly

What was fixed:

• Root cause was NOT vLLM - it was the deprecated HF_HUB_ENABLE_HF_TRANSFER=1 env var
• The script was setting this env var but hf_transfer package no longer exists
• This caused download failures that manifested as "Can't load image processor" errors
• Fix: Removed the HF_HUB_ENABLE_HF_TRANSFER=1 setting from the script

Test results (2026-01-29):

• 10/10 samples processed successfully
• Clean markdown output with proper headers and paragraphs
• Output dataset: davanstrien/lighton-ocr2-test-v4

Example usage:

hf jobs uv run --flavor a100-large \
    -s HF_TOKEN \
    https://huggingface.co/datasets/uv-scripts/ocr/raw/main/lighton-ocr2.py \
    davanstrien/ufo-ColPali output-dataset \
    --max-samples 10 --shuffle --seed 42

Model Info:

• Model: lightonai/LightOnOCR-2-1B
• Architecture: Pixtral ViT encoder + Qwen3 LLM
• Training: RLVR (Reinforcement Learning with Verifiable Rewards)
• Performance: 83.2% on OlmOCR-Bench, 42.8 pages/sec on H100

PaddleOCR-VL-1.5 (paddleocr-vl-1.5.py)

✅ Production Ready (Added 2026-01-30)

Status: Working with transformers

Note: Uses transformers backend (not vLLM) because PaddleOCR-VL only supports vLLM in server mode, which doesn't fit the single-command UV script pattern. Images are processed one at a time for stability.

Test results (2026-01-30):

• 10/10 samples processed successfully
• Processing time: ~50s per image on L4 GPU
• Output dataset: davanstrien/paddleocr-vl15-final-test

Example usage:

hf jobs uv run --flavor l4x1 \
    -s HF_TOKEN \
    https://huggingface.co/datasets/uv-scripts/ocr/raw/main/paddleocr-vl-1.5.py \
    davanstrien/ufo-ColPali output-dataset \
    --max-samples 10 --shuffle --seed 42

Task modes:

• ocr (default): General text extraction to markdown
• table: Table extraction to HTML format
• formula: Mathematical formula recognition to LaTeX
• chart: Chart and diagram analysis
• spotting: Text spotting with localization (uses higher resolution)
• seal: Seal and stamp recognition

Model Info:

• Model: PaddlePaddle/PaddleOCR-VL-1.5
• Size: 0.9B parameters (ultra-compact)
• Performance: 94.5% SOTA on OmniDocBench v1.5
• Backend: Transformers (single image processing)
• Requires: transformers>=5.0.0

Pending Development

DeepSeek-OCR-2 (Visual Causal Flow Architecture)

Status: ⏳ Waiting for vLLM upstream support

Context: DeepSeek-OCR-2 is the next generation OCR model (3B parameters) with Visual Causal Flow architecture offering improved quality. We attempted to create a UV script (deepseek-ocr2-vllm.py) but encountered a blocker.

Blocker: vLLM does not yet support DeepseekOCR2ForCausalLM architecture in the official release.

PR to Watch: 🔗 https://github.com/vllm-project/vllm/pull/33165

This PR adds DeepSeek-OCR-2 support but is currently:

• ⚠️ Open (not merged)
• Has unresolved review comments
• Pre-commit checks failing
• Issues: hardcoded parameters, device mismatch bugs, missing error handling

What's Needed:

1. PR #33165 needs to be reviewed, fixed, and merged
2. vLLM needs to release a version including the merge
3. Then we can add these dependencies to our script:

   # dependencies = [
   #     "datasets>=4.0.0",
   #     "huggingface-hub",
   #     "pillow",
   #     "vllm",
   #     "tqdm",
   #     "toolz",
   #     "torch",
   #     "addict",
   #     "matplotlib",
   # ]
   

Implementation Progress:

• ✅ Created deepseek-ocr2-vllm.py script
• ✅ Fixed dependency issues (pyarrow, datasets>=4.0.0)
• ✅ Tested script structure on HF Jobs
• ❌ Blocked: vLLM doesn't recognize architecture

Partial Implementation: The file deepseek-ocr2-vllm.py exists in this repo but is not functional until vLLM support lands. Consider it a draft.

Testing Evidence: When we ran on HF Jobs, we got:

ValidationError: Model architectures ['DeepseekOCR2ForCausalLM'] are not supported for now.
Supported architectures: [...'DeepseekOCRForCausalLM'...]

Next Steps (when PR merges):

1. Update deepseek-ocr2-vllm.py dependencies to include addict and matplotlib
2. Test on HF Jobs with small dataset (10 samples)
3. Verify output quality
4. Update README.md with DeepSeek-OCR-2 section
5. Document v1 vs v2 differences

Alternative Approaches (if urgent):

• Create transformers-based script (slower, no vLLM batching)
• Use DeepSeek's official repo setup (complex, not UV-script compatible)

Model Information:

• Model ID: deepseek-ai/DeepSeek-OCR-2
• Model Card: https://huggingface.co/deepseek-ai/DeepSeek-OCR-2
• GitHub: https://github.com/deepseek-ai/DeepSeek-OCR-2
• Parameters: 3B
• Resolution: (0-6)×768×768 + 1×1024×1024 patches
• Key improvement: Visual Causal Flow architecture

Resolution Modes (for v2):

RESOLUTION_MODES = {
    "tiny": {"base_size": 512, "image_size": 512, "crop_mode": False},
    "small": {"base_size": 640, "image_size": 640, "crop_mode": False},
    "base": {"base_size": 1024, "image_size": 768, "crop_mode": False},  # v2 optimized
    "large": {"base_size": 1280, "image_size": 1024, "crop_mode": False},
    "gundam": {"base_size": 1024, "image_size": 768, "crop_mode": True},  # v2 optimized
}

Other OCR Scripts

Nanonets OCR (nanonets-ocr.py, nanonets-ocr2.py)

✅ Both versions working

PaddleOCR-VL (paddleocr-vl.py)

✅ Working

---

Future: OCR Smoke Test Dataset

Status: Idea (noted 2026-02-12)

Build a small curated dataset (uv-scripts/ocr-smoke-test?) with ~2-5 samples from diverse sources. Purpose: fast CI-style verification that scripts still work after dep updates, without downloading full datasets.

Design goals:

• Tiny (~20-30 images total) so download is seconds not minutes
• Covers the axes that break things: document type, image quality, language, layout complexity
• Has ground truth text where possible for quality regression checks
• All permissively licensed (CC0/CC-BY preferred)

Candidate sources:

Source	What it covers	Why
NationalLibraryOfScotland/medical-history-of-british-india	Historical English, degraded scans	Has hand-corrected text column for comparison. CC0. Already tested with GLM-OCR.
davanstrien/ufo-ColPali	Mixed modern documents	Already used as our go-to test set. Varied layouts.
Something with tables	Structured data extraction	Tests --task table modes. Maybe a financial report or census page.
Something with formulas/LaTeX	Math notation	Tests --task formula. arXiv pages or textbook scans.
Something multilingual (CJK, Arabic, etc.)	Non-Latin scripts	GLM-OCR claims zh/ja/ko support. Good to verify.
Something handwritten	Handwriting recognition	Edge case that reveals model limits.

How it would work:

# Quick smoke test for any script
uv run glm-ocr.py uv-scripts/ocr-smoke-test smoke-out --max-samples 5
# Or a dedicated test runner that checks all scripts against it

Open questions:

• Build as a proper HF dataset, or just a folder of images in the repo?
• Should we include expected output for regression testing (fragile if models change)?
• Could we add a --smoke-test flag to each script that auto-uses this dataset?
• Worth adding to HF Jobs scheduled runs for ongoing monitoring?

---

Last Updated: 2026-02-12 Watch PRs:

• DeepSeek-OCR-2: https://github.com/vllm-project/vllm/pull/33165