A LoRA-fine-tuned version of Qwen2.5-0.5B specialized in Uzbek legal language. Designed to assist with legal text understanding, document processing, and Q&A in Uzbek.

📋 Model Description

This is a Parameter-Efficient Fine-Tuning (PEFT) adapter trained on top of Qwen2.5-0.5B using LoRA. It addresses the gap of low-resource NLP for the Uzbek language, particularly in the legal domain.

🎯 Intended Use Cases

Uzbek legal document Q&A
Summarization of Uzbek laws and regulations
Legal chatbot for Uzbek-speaking users
Foundation for further fine-tuning on Uzbek legal tasks
Research on low-resource language models

🚀 Quick Start

Install required libraries:

pip install transformers peft torch

Load the model with PEFT:

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "Qwen/Qwen2.5-0.5B"
adapter = "AsrorAsr/qwen2.5-0.5b-uzbek-legal-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(base_model)
model = PeftModel.from_pretrained(model, adapter)

prompt = "Mehnat shartnomasi nima?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=150)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

📊 Training Details

Base model: Qwen2.5-0.5B (494M parameters)
Method: LoRA (Low-Rank Adaptation)
Framework: PEFT + Transformers + PyTorch
Training data: Uzbek legal documents
Related dataset: lexuz-legal-documents

📚 Dataset

Trained using the author's curated Uzbek legal datasets:

AsrorAsr/lexuz-legal-documents (27.1k documents)
AsrorAsr/uzbek-legal-dataset (1.22k documents)

⚠️ Limitations & Bias

Small base model (0.5B) — limited reasoning capability
Not a substitute for professional legal advice
May produce inaccurate or outdated information
Best used as an assistant tool, not a final authority
Reflects biases present in training documents

🌐 Why Uzbek Legal AI Matters

Uzbek is a low-resource language in NLP. This model contributes to bridging that gap by enabling AI tools for Uzbek-speaking professionals, students, and citizens working with legal content.

📜 License

Apache 2.0 — free for commercial and personal use.

👤 Author

Asror Akmalov | AI/ML Engineer
🔗 GitHub
🔗 Hugging Face

Open to freelance projects and collaboration on low-resource NLP and Uzbek language AI.

📬 Citation

@misc{akmalov2025uzbeklegalqwen,
  author = {Akmalov, Asror},
  title = {Qwen2.5-0.5B Uzbek Legal LoRA},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/AsrorAsr/qwen2.5-0.5b-uzbek-legal-lora}
}