Callisto-OCR3-2B-Instruct [ VL / OCR ]

The Callisto-OCR3-2B-Instruct model is a fine-tuned version of Qwen2-VL-2B-Instruct, specifically optimized for messy handwriting recognition, Optical Character Recognition (OCR), English language understanding, and math problem solving with LaTeX formatting. This model integrates a conversational approach with visual and textual understanding to handle multi-modal tasks effectively.

Key Enhancements:

• SoTA understanding of images of various resolution & ratio: Callisto-OCR3 achieves state-of-the-art performance on visual understanding benchmarks, including MathVista, DocVQA, RealWorldQA, MTVQA, etc.

• Enhanced Handwriting OCR: Optimized for recognizing and interpreting messy handwriting with high accuracy, making it ideal for digitizing handwritten documents and notes.

• Understanding videos of 20min+: Callisto-OCR3 can process long videos, enabling high-quality video-based question answering, transcription, and content generation.

• Agent that can operate your mobiles, robots, etc.: With advanced reasoning and decision-making, Callisto-OCR3 can be integrated with mobile phones, robots, and other devices to perform automated tasks based on visual and textual input.

• Multilingual Support: Besides English and Chinese, Callisto-OCR3 supports text recognition inside images in multiple languages, including European languages, Japanese, Korean, Arabic, and Vietnamese.

How to Use

from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info

# Load the model on the available device(s)
model = Qwen2VLForConditionalGeneration.from_pretrained(
    "prithivMLmods/Callisto-OCR3-2B-Instruct", torch_dtype="auto", device_map="auto"
)

# Enable flash_attention_2 for better acceleration and memory optimization
# model = Qwen2VLForConditionalGeneration.from_pretrained(
#     "prithivMLmods/Callisto-OCR3-2B-Instruct",
#     torch_dtype=torch.bfloat16,
#     attn_implementation="flash_attention_2",
#     device_map="auto",
# )

# Default processor
processor = AutoProcessor.from_pretrained("prithivMLmods/Callisto-OCR3-2B-Instruct")

# Customize visual token range for speed-memory balance
# min_pixels = 256*28*28
# max_pixels = 1280*28*28
# processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct", min_pixels=min_pixels, max_pixels=max_pixels)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
            },
            {"type": "text", "text": "Recognize the handwriting in this image."},
        ],
    }
]

# Preparation for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Inference: Generate the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

Buffering Output

buffer = ""
for new_text in streamer:
    buffer += new_text
    # Remove <|im_end|> or similar tokens from the output
    buffer = buffer.replace("<|im_end|>", "")
    yield buffer

Key Features

1. Advanced Handwriting OCR:

   • Excels at recognizing and transcribing messy and cursive handwriting into digital text with high accuracy.

2. Vision-Language Integration:

   • Combines image understanding with natural language processing to convert images into text.

3. Optical Character Recognition (OCR):

   • Extracts and processes textual information from images with precision.

4. Math and LaTeX Support:

   • Solves math problems and outputs equations in LaTeX format.

5. Conversational Capabilities:

   • Designed to handle multi-turn interactions, providing context-aware responses.

6. Image-Text-to-Text Generation:

   • Inputs can include images, text, or a combination, and the model generates descriptive or problem-solving text.