AinBondhu is a fine-tuned LoRA adapter based on Qwen/Qwen2.5-7B-Instruct, specifically trained to act as a legal assistant for Bangladeshi law. It understands both Bengali and English, providing accurate guidance on various legal scenarios, laws, and constitutional matters in Bangladesh.

🚀 Quick Start (Inference in Python)

To seamlessly use this model, load the base model (Qwen/Qwen2.5-7B-Instruct) and then apply these LoRA adapters using peft and transformers.

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel

# Define models
LORA_MODEL = "himu1780/ainbondhu-legal-lora"
BASE_MODEL = "Qwen/Qwen2.5-7B-Instruct" 

# 4-bit Quantization Config (Great for free Colab T4 GPUs)
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.float16,
)

# Load Tokenizer & Base Model
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    quantization_config=bnb_config,
    device_map="auto",
)

# Apply AinBondhu LoRA Adapters
model = PeftModel.from_pretrained(base_model, LORA_MODEL)
model.eval()

# Chat with AinBondhu
prompt = "বাংলাদেশে সাইবার বুলিং এর শাস্তি কি? বিস্তারিত বল।"
messages = [
    {"role": "system", "content": "You are AinBondhu, a helpful legal AI assistant for Bangladesh."},
    {"role": "user", "content": prompt},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.2, # Keep low for factual accuracy
        top_p=0.9,
        pad_token_id=tokenizer.eos_token_id,
    )

print("AinBondhu says:
" + tokenizer.decode(out[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True))

📊 Training Details

Base Model: unsloth/qwen2.5-7b-instruct-bnb-4bit (Qwen 2.5 7B Instruct)
Dataset Size: 142,552 high-quality legal conversations
Epochs: 1
Training Method: Supervised Fine-Tuning (SFT) with LoRA via unsloth and trl.

⚠️ Disclaimer

This AI model is designed for educational and informational purposes only. It does not provide professional legal advice. Always consult a qualified lawyer or legal professional in Bangladesh for actual legal proceedings or advice.