Patram-7B-Instruct

Patram-7B-Instruct by BharatGen is a 7B parameter vision-language model trained from scratch for visual document understanding. As India’s first document foundation model, it is built to tackle complex document analysis. The model was trained on a carefully curated instruction-tuned dataset, combining diverse public and custom synthetic data designed to support a broad spectrum of document understanding tasks.

Model Overview

• Architecture: Vision Transformer (ViT) + MLP projector + OLMo-7B LLM
• Training Data: BharatDocs-v1, a dataset of diverse Indian documents + Other Open Source Document Datasets
• Supported I/O Formats: The model currently accepts English-language instructions and image files (e.g., PNG, JPEG) as input. The output is provided in text format.
• Language: English (Indian language support upcoming)
• License: Apache 2.0

Usage Examples

Use the transformers library.

import torch
from transformers import AutoProcessor, AutoModelForCausalLM, GenerationConfig
from PIL import Image
import requests

# Model ID and device setup
model_id = "bharatgenai/patram-7b-instruct"
device = "cuda" if torch.cuda.is_available() else "cpu"

# Load processor and model
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True
).to(device)

def get_patram_response(image_path_or_url, question):
    try:
        # Load image
        if image_path_or_url.startswith("http"):
            image = Image.open(requests.get(image_path_or_url, stream=True).raw).convert("RGB")
        else:
            image = Image.open(image_path_or_url).convert("RGB")
    except Exception as e:
        print(f"Error loading image: {e}")
        return None

    # Format the prompt as expected
    prompt = f"Question: {question} Answer based on the image."

    try:
        # Preprocess image and text using the processor
        inputs = processor.process(images=[image], text=prompt)
        inputs = {k: v.to(device).unsqueeze(0) for k, v in inputs.items()}

        # Generate output using model's generate_from_batch method (Patram-specific)
        output = model.generate_from_batch(
            inputs,
            GenerationConfig(max_new_tokens=200, stop_strings="<|endoftext|>"),
            tokenizer=processor.tokenizer
        )

        # Extract generated tokens (excluding input tokens) and decode
        generated_tokens = output[0, inputs['input_ids'].size(1):]
        response = processor.tokenizer.decode(generated_tokens, skip_special_tokens=True).strip()
        return response
    except Exception as e:
        print(f"Error during inference: {e}")
        return None

# Example usage:
# image_input = "https://knowscope.in/wp-content/uploads/2025/05/cghd-nag.png"
# question = "Who issued this notice?"
# answer = get_patram_response(image_input, question)
# if answer:
#     print("Answer:", answer)

Evaluations

We evaluated Patram-7B-Instruct alongside other vision-language models (VLMs) in the 7B–9B parameter range across multiple public document benchmarks.

Benchmarks: DocVQA, VisualMRC, Patram-Bench

Patram-Bench is an in-house benchmark designed for Indic Document VQA.

Metric: G-Eval (LLM-as-a-judge)

Model	Overall	DocVQA	Patram-Bench	VisualMRC
claude-3.7-sonnet	0.8830	0.8480	0.8857	0.8830
Qwen2.5-VL-7B-Instruct	0.8759	0.8722	0.6816	0.9169
gemma-3-12b-it	0.8556	0.8451	0.6349	0.9069
patram-7b-instruct	0.8331	0.8550	0.6515	0.8510
InternVL3-9B	0.7865	0.8681	0.6888	0.7405
deepseek-vl2	0.7581	0.8739	0.5089	0.7144

*Note: The benchmarked results reflect the API variant.

Citation

@online{BharatGenPatramLaunch2025,
  author    = {{BharatGen Team}},
  title     = {BharatGen Unveils Patram: India's Pioneering Vision-Language Foundation Model for Document Intelligence},
  year      = {2025},
  url       = {https://bharatgen.com/blog/patram-launch},
  urldate   = {2025-06-02}
}

Resources

• Model: huggingface.co/bharatgenai/patram-7b-instruct
• Project Page: bharatgen.com/patram
• Blog: bharatgen.com/blog/patram-launch

Authors

• Principal Investigators: Prof. Ravi Kiran Sarvadevabhatla, Prof. Ganesh Ramakrishnan
• Contributors: BharatGen Team

Contact

• Contact Form
• Hugging Face Community Tab