For text only LLMs: ./llama.cpp/llama-cli -hf smartytrios/docintel_ocr_llama_3_2_gguf --jinja
For multimodal models: ./llama.cpp/llama-mtmd-cli -hf smartytrios/docintel_ocr_llama_3_2_gguf --jinja

Available Model files:

Llama-3.2-1B-Instruct.Q4_K_M.gguf

Ollama

An Ollama Modelfile is included for easy deployment. This was trained 2x faster with Unsloth

tags:

gguf
llama.cpp
unsloth
ocr
document-intelligence
json-extraction license: mit datasets:
smartytrios/document_data_extractor language:
en base_model:
unsloth/Llama-3.2-1B-Instruct-bnb-4bit pipeline_tag: text-generation library_name: transformers

docintel_ocr_llama_3_2_gguf : GGUF Optimized

This model is a fine-tuned version of Llama-3.2-1B-Instruct, specialized for Document Intelligence and OCR-to-JSON extraction. It was trained using the Unsloth library to optimize memory efficiency and training speed, then exported to GGUF format for local deployment.

Model Description

The primary objective of this model is to transform unstructured text generated by Optical Character Recognition (OCR) engines into structured, machine-readable JSON formats. It is specifically tuned to handle noise, line breaks (\n), and misalignments common in raw OCR data.

Architecture: Llama 3.2 (1B Parameters)
Quantization: Q4_K_M (4-bit Medium)
Specialization: Invoice/Receipt data extraction, medical bill parsing, and form field mapping.
Fine-tuning Method: QLoRA (Rank: 16)

🚀 Usage Guide

1. Local Inference with llama.cpp

For the best performance on Windows, Mac, or Linux using llama.cpp, use the following command:

./llama-cli -hf smartytrios/docintel_ocr_llama_3_2_gguf --jinja -p "### OCR:\n[PASTE YOUR OCR TEXT HERE]\n### JSON:"