stanford-oval/churro-3B
CHURRO-3B is a 3 billion parameter open-weight vision-language model (VLM) developed by Stanford OVAL, specifically designed for high-accuracy, low-cost historical document transcription. Built upon the Qwen2.5-VL-3B-Instruct architecture, it excels at recognizing handwritten and printed text across 22 centuries and 46 language clusters, including historical and dead languages. This model offers a significant cost advantage, achieving higher accuracy than much larger commercial and open-weight VLMs like Gemini 2.5 Pro at 15.5x lower cost, making it ideal for historical text recognition tasks.
Loading preview...
CHURRO-3B: High-Accuracy, Low-Cost Historical Text Recognition
CHURRO-3B is a 3-billion parameter open-weight vision-language model (VLM) developed by Stanford OVAL, specifically engineered for the transcription of historical documents. It is built upon the Qwen2.5-VL-3B-Instruct base model and features a context length of 32768 tokens.
Key Capabilities
- Historical Text Recognition: Specializes in transcribing both handwritten and printed text from historical documents.
- Broad Linguistic and Temporal Coverage: Capable of recognizing text across 22 centuries and 46 language clusters, including historical and dead languages.
- Cost-Efficiency: Achieves higher accuracy than significantly larger commercial and open-weight VLMs (e.g., Gemini 2.5 Pro) at a substantially lower cost (15.5x lower).
- Extensive Training Data: Trained on CHURRO-DS, a curated dataset comprising approximately 100,000 pages from 155 historical collections.
Good For
- Researchers and institutions working with historical archives and documents.
- Applications requiring accurate and cost-effective transcription of ancient or historical texts.
- Projects focused on digitizing and making historical manuscripts and records searchable.