170 Downloads Updated 1 week ago
Chandra-OCR-2 from Datalab is a state-of-the-art OCR model that outputs structured markdown, HTML, or JSON while preserving precise layout information from images and PDFs across 90+ languages. there is another of this in Ollama but unpatched [no vision?]
ollama run fredrezones55/chandra-ocr-2:patch
Details
Updated 1 week ago
1 week ago
8714a4782ea6 · 5.8GB ·
Readme
Seems to have been a requested model though the Ollama issues page, what is the point of having this model if there is no vision capability? We just needed to correct this mistake.
before I forget: Happy Easter!
the GGUF model was sourced from: https://huggingface.co/prithivMLmods/chandra-ocr-2-GGUF the original fine-tune was from: https://huggingface.co/datalab-to/chandra-ocr-2
As an experiment: I’ll start this 4B model with Q8 quant paired with F16 vision. [this config seems to mostly fit in a 8GB vram pascal gpu.]
Noting that the base model is Qwen3.5:4B and all it’s limitations with Ollama, but this model has vision fully working otherwise that breaks the point of a OCR vision based model 🤣.
you could just give the patched model no prompt and just the image; and it will begin to OCR unprompted.
the model has likely been trained thoroughly with datalab-to’s OCR model harness with the quick start recommentation of:
pip install chandra-ocr
# With vLLM (recommended, easy install)
chandra_vllm
chandra input.pdf ./output
# With HuggingFace (requires torch)
pip install chandra-ocr[hf]
chandra input.pdf ./output --method hf
# In particular this ollama model, we can use vllm
VLLM_API_BASE=http://localhost:11434/v1 VLLM_MODEL_NAME=fredrezones55/chandra-ocr-2:patch chandra --method vllm input output
the patch model is my attempts to constrain the base model so it will stop thinking and breaking the chandra program. [he could have used an instruction model or something] {or perhaps I have not done enough research}
issues could be a capped text generation where you might need to set the MAX_OUTPUT_TOKENS environment variable.
Multilingual Benchmark (43 Languages)
The table below covers the 43 most common languages, benchmarked across multiple models. For a comprehensive evaluation across 90 languages (Chandra 2 vs Gemini 2.5 Flash only), see the full 90-language benchmark.
| Language | Datalab API | Chandra 2 | Chandra 1 | Gemini 2.5 Flash | GPT-5 Mini |
|---|---|---|---|---|---|
| ar | 67.6% | 68.4% | 34.0% | 84.4% | 55.6% |
| bn | 85.1% | 72.8% | 45.6% | 55.3% | 23.3% |
| ca | 88.7% | 85.1% | 84.2% | 88.0% | 78.5% |
| cs | 88.2% | 85.3% | 84.7% | 79.1% | 78.8% |
| da | 90.1% | 91.1% | 88.4% | 86.0% | 87.7% |
| de | 93.8% | 94.8% | 83.0% | 88.3% | 93.8% |
| el | 89.9% | 85.6% | 85.5% | 83.5% | 82.4% |
| es | 91.8% | 89.3% | 88.7% | 86.8% | 97.1% |
| fa | 82.2% | 75.1% | 69.6% | 61.8% | 56.4% |
| fi | 85.7% | 83.4% | 78.4% | 86.0% | 84.7% |
| fr | 93.3% | 93.7% | 89.6% | 86.1% | 91.1% |
| gu | 73.8% | 70.8% | 44.6% | 47.6% | 11.5% |
| he | 76.4% | 70.4% | 38.9% | 50.9% | 22.3% |
| hi | 80.5% | 78.4% | 70.2% | 82.7% | 41.0% |
| hr | 93.4% | 90.1% | 85.9% | 88.2% | 81.3% |
| hu | 88.1% | 82.1% | 82.5% | 84.5% | 84.8% |
| id | 91.3% | 91.6% | 86.7% | 88.3% | 89.7% |
| it | 94.4% | 94.1% | 89.1% | 85.7% | 91.6% |
| ja | 87.3% | 86.9% | 85.4% | 80.0% | 76.1% |
| jv | 87.5% | 73.2% | 85.1% | 80.4% | 69.6% |
| kn | 70.0% | 63.2% | 20.6% | 24.5% | 10.1% |
| ko | 89.1% | 81.5% | 82.3% | 84.8% | 78.4% |
| la | 78.0% | 73.8% | 55.9% | 70.5% | 54.6% |
| ml | 72.4% | 64.3% | 18.1% | 23.8% | 11.9% |
| mr | 80.8% | 75.0% | 57.0% | 69.7% | 20.9% |
| nl | 90.0% | 88.6% | 85.3% | 87.5% | 83.8% |
| no | 89.2% | 90.3% | 85.5% | 87.8% | 87.4% |
| pl | 93.8% | 91.5% | 83.9% | 89.7% | 90.4% |
| pt | 97.0% | 95.2% | 84.3% | 89.4% | 90.8% |
| ro | 86.2% | 84.5% | 82.1% | 76.1% | 77.3% |
| ru | 88.8% | 85.5% | 88.7% | 82.8% | 72.2% |
| sa | 57.5% | 51.1% | 33.6% | 44.6% | 12.5% |
| sr | 95.3% | 90.3% | 82.3% | 89.7% | 83.0% |
| sv | 91.9% | 92.8% | 82.1% | 91.1% | 92.1% |
| ta | 82.9% | 77.7% | 50.8% | 53.9% | 8.1% |
| te | 69.4% | 58.6% | 19.5% | 33.3% | 9.9% |
| th | 71.6% | 62.6% | 47.0% | 66.7% | 53.8% |
| tr | 88.9% | 84.1% | 68.1% | 84.1% | 78.2% |
| uk | 93.1% | 91.0% | 88.5% | 87.9% | 81.9% |
| ur | 54.1% | 43.2% | 28.1% | 57.6% | 16.9% |
| vi | 85.0% | 80.4% | 81.6% | 89.5% | 83.6% |
| zh | 87.8% | 88.7% | 88.3% | 70.0% | 70.4% |
| Average | 80.4% | 77.8% | 69.4% | 67.6% | 60.5% |
Full 90-Language Benchmark
We also have a more comprehensive evaluation covering 90 languages, comparing Chandra 2 against Gemini 2.5 Flash. The average scores are lower than the 43-language table above because this includes many lower-resource languages. Chandra 2 averages 72.7% vs Gemini 2.5 Flash at 60.8%.
See the full 90-language results.