Patram-7B-Instruct

Patram by BharatGen is a family of 7B parameter vision-language models trained from scratch for visual document understanding. This README covers Patram-7B-Instruct (v1.0), Patram1.1-7B-Instruct, and Patram1.2-7B-Instruct, released as revisions of the same bharatgenai/patram-7b-instruct repository — pick a revision below and run any of them.

News 📢

• [2026-02-17] 🚀 Released Patram1.2-7B-Instruct — adds Chain-of-thought VQA, Text Extraction, Chart Understanding, Table Understanding, and Flowchart Understanding on top of v1.1, trained on BharatDocs-v1.2. Evaluated across Text-Rich Document VQA, Visually-Rich Document VQA, TableVQA, and Indian benchmarks.
• [2025-08-08] 🚀 Released Patram1.1-7B-Instruct — adds Multi-turn VQA and Key-value extraction on top of v1.0, trained on BharatDocs-v1.1. Evaluated across Text-Rich Document VQA, Visually-Rich Document VQA, and Indian benchmarks.
• [2025-06-02] 🚀 Released Patram-7B-Instruct — India's first document foundation model for visual document understanding, trained on BharatDocs-v1.

Index

1. News
2. Introduction
3. Comparison
4. Benchmarks
   • Patram1.2 — Text-Rich Document VQA
   • Patram1.2 — Visually-Rich Document VQA
   • Patram1.2 — TableVQA
   • Patram1.2 — Indian Benchmarks & Overall Average
5. Inference
   • Environment
   • Hugging Face
   • Footnote — Patram1.2 prompting modes
6. Examples
7. Authors
8. Citation
9. Contact

Introduction

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. Each release is 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.

• Patram-7B-Instruct (v1.0) is the original release.
• Patram1.1-7B-Instruct builds on it with a richer training set (BharatDocs-v1.1) and adds multi-turn and key-value-extraction capabilities.
• Patram1.2-7B-Instruct builds further (BharatDocs-v1.2), adding chain-of-thought reasoning, text extraction, and chart/table/flowchart understanding.

Comparison

Benchmarks

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

Overall Avg. is the average of the Text-Rich Document VQA, Visually-Rich Document VQA, TableVQA, and Patram-Bench-VQA-en bucket averages above.

Inference

All three versions live in the same Hugging Face repository, bharatgenai/patram-7b-instruct, as separate revisions — patram1.0, patram1.1, and patram1.2. Set REVISION accordingly in the snippet below.

Environment

All three revisions use the same environment

torch==2.7.0
torchvision==0.22.0
transformers==4.50.3
accelerate==0.26.0
einops==0.8.1

import base64
import io
import urllib.request

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

MODEL_ID = "bharatgenai/patram-7b-instruct"
REVISION = "patram1.1"  # "patram1.0", "patram1.1", or "patram1.2"
HF_TOKEN = "hf_........"  # Insert HF token here
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

IMAGE_URL = "https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcQzcRxT-eX_C8Kahj9-DvCF1t1gzHctM7SIFOasQkCboQ&s=10"  # Insert image URL here
IMAGE_B64 = ""                                  # ...or a base64-encoded image instead


def load_image(url=None, b64=None):
    """Load a document image from a URL or from a base64 string."""
    if b64:
        # tolerates a leading "data:image/png;base64," prefix
        data = base64.b64decode(b64.split(",", 1)[-1])
    elif url:
        with urllib.request.urlopen(url) as response:
            data = response.read()
    else:
        raise ValueError("Provide either url= or b64=")
    return Image.open(io.BytesIO(data)).convert("RGB")


image = load_image(url=IMAGE_URL)
# image = load_image(b64=IMAGE_B64)

processor = AutoProcessor.from_pretrained(
    MODEL_ID, revision=REVISION, trust_remote_code=True, token=HF_TOKEN
)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    revision=REVISION,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    token=HF_TOKEN,
).to(DEVICE)
model.eval()

prompt = "Question: What is this document about? Answer based on the image."
inputs = processor.process(images=[image], text=prompt)
# float tensors (the image patches) must match the model dtype; integer ids must not be cast
inputs = {
    k: (v.to(DEVICE).to(model.dtype) if torch.is_floating_point(v) else v.to(DEVICE)).unsqueeze(0)
    for k, v in inputs.items()
}

with torch.autocast(device_type=DEVICE, enabled=True, dtype=torch.bfloat16):
    output = model.generate_from_batch(
        inputs,
        GenerationConfig(max_new_tokens=256, do_sample=False, stop_strings="<|endoftext|>"),
        tokenizer=processor.tokenizer,
    )

generated = output[0, inputs["input_ids"].size(1):]
print(processor.tokenizer.decode(generated, skip_special_tokens=True))

Patram1.2 only — supports two prompting modes via a MODE flag, direct and cot:

MODE = "cot"  # "direct" or "cot"
PROMPT_TEMPLATES = {
    "direct": "Question: {question} Answer based on the image.",
    "cot": "Question: {question} Think step by step, then answer based on the image.",
}
MAX_NEW_TOKENS = {"direct": 2048, "cot": 4096}

question = "What is the total amount on this invoice?"
prompt = PROMPT_TEMPLATES[MODE].format(question=question)
inputs = processor.process(images=[image], text=prompt)  # image = PIL.Image.open(...).convert("RGB")

Note: If you're trying this on an Apple Silicon (M1/M2/M3/M4/...) chip, please follow the official documentation by PyTorch and Hugging Face for installing dependencies:

• PyTorch's official guide on installation (macOS)
• Hugging Face Transformers performance tips

Footnote — Patram1.2 prompting modes

Patram1.2 supports two prompting modes (direct and cot — see Hugging Face above for how to run each). Benchmark scores above were evaluated with the following mode per benchmark:

Examples

Examples below are from Patram1.1-7B-Instruct.

Examples below are from Patram1.2-7B-Instruct.

Authors

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

Citation

@misc{bharatgen_patram_2025,
  title        = {Patram: India's Pioneering Vision-Language Foundation Model for Document Intelligence},
  author       = {{BharatGen Team}},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/bharatgenai/patram-7b-instruct}},
  note         = {Accessed: 2025-06-02}
}

@misc{bharatgen_patram1_1_2025,
  title        = {Patram1.1: India's Pioneering Vision-Language Foundation Model for Document Intelligence},
  author       = {{BharatGen Team}},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/bharatgenai/patram-7b-instruct}},
  note         = {Accessed: 2025-08-08}
}

@misc{bharatgen_patram1_2_2026,
  title        = {Patram1.2: India's Pioneering Vision-Language Foundation Model for Document Intelligence},
  author       = {{BharatGen Team}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/bharatgenai/patram-7b-instruct}},
  note         = {Accessed: 2026-02-17}
}

Contact

• Contact Form
• Hugging Face Community Tab