Prarabdha/pixtral-12b-240910-hf

Pixtral-12B-2409 - HuggingFace Transformers Compatible Weights

Model Overview

This repository contains the HuggingFace Transformers compatible weights for the Pixtral-12B-2409 multimodal model. The weights have been converted to ensure seamless integration with the Hugging Face Transformers library, allowing easy loading and usage in your projects.

Model Details

• Original Model: Pixtral-12B-2409 by Mistral AI
• Model Type: Multimodal Language Model
• Parameters: 12B parameters + 400M parameter vision encoder
• Sequence Length: 128k tokens
• License: Apache 2.0

Key Features

• Natively multimodal, trained with interleaved image and text data
• Supports variable image sizes
• Leading performance in its weight class on multimodal tasks
• Maintains state-of-the-art performance on text-only benchmarks

Conversion Details

This repository provides the original Pixtral model weights converted to be fully compatible with the HuggingFace Transformers library. The conversion process ensures:

• Seamless loading using from_pretrained()
• Full compatibility with HuggingFace Transformers pipeline
• No modifications to the original model weights or architecture

Installation

You can install the model using the Transformers library:

from transformers import AutoProcessor, AutoModelForImageTextToText
import torch

processor = AutoProcessor.from_pretrained("Prarabdha/pixtral-12b-240910-hf")
model = AutoModelForImageTextToText.from_pretrained("Prarabdha/pixtral-12b-240910-hf", torch_dtype=torch.float16, device_map="auto")

Example Usage

from PIL import Image
import requests

# Load an image
url = "https://example.com/sample-image.jpg"
image = Image.open(requests.get(url, stream=True).raw)

# Prepare conversation
conversation = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": "What is shown in this image?"},
        ],
    }
]

# Process and generate
prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
inputs = processor(images=[image], text=prompt, return_tensors="pt")
generate_ids = model.generate(**inputs, max_new_tokens=30)
response = processor.batch_decode(generate_ids, skip_special_tokens=True)

Performance Benchmarks

Multimodal Benchmarks

Benchmark	Pixtral 12B	Qwen2 7B VL	LLaVA-OV 7B	Phi-3 Vision
MMMU (CoT)	52.5	47.6	45.1	40.3
Mathvista (CoT)	58.0	54.4	36.1	36.4
ChartQA (CoT)	81.8	38.6	67.1	72.0

(Full benchmark details available in the original model card)

Acknowledgements

A huge thank you to the Mistral team for creating and releasing the original Pixtral model.

Citation

If you use this model, please cite the original Mistral AI research.

License

This model is distributed under the Apache 2.0 License.

Original Model Card

For more comprehensive details, please refer to the original Mistral model card.