Fine-tuned Qwen/Qwen2.5-1.5B-Instruct on Indian government scheme Q&A in English and Hindi. Part of the Yojana Sahayak project — a multilingual voice assistant for government welfare schemes.

Eval Results

Metric	Score
Train loss	0.3375
Eval loss	0.2073
Perplexity	1.15
ROUGE-1	0.1762
ROUGE-2	0.1501
ROUGE-L	0.1754
Train samples	10,000

Training

Base model: Qwen/Qwen2.5-1.5B-Instruct
Method: QLoRA (4-bit NF4 + LoRA r=16, alpha=32)
Dataset: Subh24ai/yojana-sahayak-instruct (10,000 train samples)
Hardware: Kaggle T4 GPU (55.2 min)

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
model = PeftModel.from_pretrained(base, "Subh24ai/yojana-sahayak-qwen2.5-1.5b-qlora")
tokenizer = AutoTokenizer.from_pretrained("Subh24ai/yojana-sahayak-qwen2.5-1.5b-qlora")

messages = [
    {"role": "system", "content": "You are Yojana Sahayak..."},
    {"role": "user",   "content": "PM Kisan ke liye kaun eligible hai?"}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(output[0], skip_special_tokens=True))