metadata

base_model: unsloth/Llama-3.2-11B-Vision-Instruct
tags:
  - text-generation-inference
  - transformers
  - unsloth
  - mllama
  - vision-language
  - document-understanding
  - data-extraction
license: apache-2.0
language:
  - en
library_name: transformers
model-index:
  - name: PixelParse_AI
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: IFEval (0-Shot)
          type: wis-k/instruction-following-eval
          split: train
          args:
            num_few_shot: 0
        metrics:
          - type: inst_level_strict_acc and prompt_level_strict_acc
            value: 43.83
            name: averaged accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Daemontatox%2FPixelParse_AI
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: BBH (3-Shot)
          type: SaylorTwift/bbh
          split: test
          args:
            num_few_shot: 3
        metrics:
          - type: acc_norm
            value: 29.03
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Daemontatox%2FPixelParse_AI
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MATH Lvl 5 (4-Shot)
          type: lighteval/MATH-Hard
          split: test
          args:
            num_few_shot: 4
        metrics:
          - type: exact_match
            value: 14.43
            name: exact match
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Daemontatox%2FPixelParse_AI
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GPQA (0-shot)
          type: Idavidrein/gpqa
          split: train
          args:
            num_few_shot: 0
        metrics:
          - type: acc_norm
            value: 9.84
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Daemontatox%2FPixelParse_AI
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MuSR (0-shot)
          type: TAUR-Lab/MuSR
          args:
            num_few_shot: 0
        metrics:
          - type: acc_norm
            value: 9.25
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Daemontatox%2FPixelParse_AI
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU-PRO (5-shot)
          type: TIGER-Lab/MMLU-Pro
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 30.87
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Daemontatox%2FPixelParse_AI
          name: Open LLM Leaderboard

Vision-Language Model for Document Data Extraction

Developed by: Daemontatox
License: apache-2.0
Finetuned from model: unsloth/Llama-3.2-11B-Vision-Instruct

Overview

This Vision-Language Model (VLM) is purpose-built for extracting structured and unstructured data from various types of documents, including but not limited to:

Invoices
Timesheets
Contracts
Forms
Receipts

By utilizing advanced multimodal learning capabilities, this model understands both text and visual layout features, enabling it to parse even complex document structures.

Key Features

Accurate Data Extraction:
- Automatically detects and extracts key fields such as dates, names, amounts, itemized details, and more.
- Outputs data in clean and well-structured JSON format.
Robust Multimodal Understanding:
- Processes both text and visual layout elements (tables, headers, footers).
- Adapts to various document formats and layouts without additional fine-tuning.
Optimized Performance:
- Fine-tuned using Unsloth, enabling 2x faster training.
- Employs Hugging Face’s TRL library for parameter-efficient fine-tuning.
Flexible Deployment:
- Compatible with a wide range of platforms for integration into document processing pipelines.
- Optimized for inference on GPUs and high-performance environments.

Use Cases

Enterprise Automation: Automate data entry and document processing tasks in finance, HR, and legal domains.
E-invoicing: Extract critical invoice details for seamless integration with ERP systems.
Compliance: Extract and structure data for auditing and regulatory compliance reporting.

Training and Fine-Tuning

The fine-tuning process leveraged Unsloth's efficiency optimizations, reducing training time while maintaining high accuracy. The model was trained on a diverse dataset of scanned documents and synthetic examples to ensure robustness across real-world scenarios.

Acknowledgments

This model was fine-tuned using the powerful capabilities of the Unsloth framework, which significantly accelerates the training of large models.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here! Summarized results can be found here!

Metric	Value (%)
Average	22.87
IFEval (0-Shot)	43.83
BBH (3-Shot)	29.03
MATH Lvl 5 (4-Shot)	14.43
GPQA (0-shot)	9.84
MuSR (0-shot)	9.25
MMLU-PRO (5-shot)	30.87