You can deploy this model and start earning money today!

4maan4hmad/
Mistral-finetuned-productidentifier-v2.0

Name: Mistral-finetuned-productidentifier-v2.0
Availability: InStock

gguf

mistral

endpoints_compatible

Views

No views yet

🧠 4maan4hmad / mistral-7b-ocr-cleaner

Fine-tuned Mistral 7B model optimized for cleaning and structuring OCR outputs

👨‍💻 Developed by

Amaan Ahmad
🔗 LinkedIn | GitHub | Hugging Face
📜 License: Apache-2.0
🧩 Base Model: unsloth/mistral-7b-instruct-v0.3-bnb-4bit

📘 Model Overview

mistral-7b-ocr-cleaner is a fine-tuned version of Mistral 7B, trained for structured text extraction and noise cleaning from OCR outputs.
It efficiently converts unstructured, messy text (e.g., from scanned packaging) into well-formatted JSON with standardized fields.

This model leverages Unsloth and Hugging Face's TRL library for adapter-based instruction tuning and memory-efficient training.

🎯 Intended Use

Ideal for:

🧾 Structured information extraction from noisy OCR text
🏷️ Product data cleaning for retail, logistics, and e-commerce pipelines
💾 Transforming free-form OCR text into consistent JSON with fields like:
- company_name
- item_name
- category
- storage_recommendation

💡 Example

🔍 Input


DOTEKEY DOT&KEY DOTEKE DOT&SKEY SLEMISHE MA E CEV N M HM DOTEKEY GELFACEWASH CALMING CICA DOTCKEY LE WASH BLEMISH RING DOT&KEY GEL FACE WASH DOTCTKEY SLEM XRE m for a clear and refreeshed face

🧾 Output

json

{
  "company_name": "DOT and KEY",
  "item_name": "Face wash",
  "category": "Self Hygiene",
  "storage_recommendation": "Warehouse Shelf"
}

🧩 Training Details

Attribute	Details
Trainer	`SFTTrainer` (Hugging Face Transformers)
Framework	PyTorch + TRL + Unsloth
Fine-tuning Type	LoRA adapter (merged into 4-bit base)
Dataset	Custom OCR pairs (raw text → structured JSON)
Precision	4-bit quantized (bnb-4bit)
Base Model	`unsloth/mistral-7b-instruct-v0.3-bnb-4bit`

🧱 Tech Stack

Libraries: Transformers, TRL, Unsloth, PEFT, BitsAndBytes
Training Hardware: A100 / RTX GPU (LoRA fine-tuning)
Format: Adapter merged for lightweight inference

❤️ Made With

📫 Connect

If you’re working on OCR data extraction, structured LLM pipelines, or AI agents for e-commerce, feel free to connect!

🔗 LinkedIn: linkedin.com/in/AmaanAhmad 💻 GitHub: github.com/4maan4hmad

4maan4hmad/Mistral-finetuned-productidentifier-v2.0