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

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-8B", torch_dtype=torch.float16, device_map="auto"
)
model = PeftModel.from_pretrained(base, "akarki15/nepali-rapper-lora")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")

messages = [
    {"role": "system", "content": (
        "Timi euta Nepali rapper ho — street bata aako, bars haru fire chha, "
        "rhymes tight chha. Timi Nepali slang, hip-hop lingo, ra Devanagari mix "
        "garera bolchau. Timi verse lekchau, freestyle garchau, ra rapper jastai "
        "kura garchau. Dherai swag, dherai attitude, tara real ra raw."
    )},
    {"role": "user", "content": "Euta verse lekha Nepal ko baare ma"},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Training Details

Base model: unsloth/Qwen3-8B-unsloth-bnb-4bit (4-bit quantized)
Method: LoRA (r=16, alpha=16, targeting all linear projections)
Data: ~50 multi-turn conversations in ShareGPT format
Topics: Verse/freestyle generation, diss tracks, battle rap, Nepal-themed raps, casual rapper chat
Languages: Mixed Nepali (Devanagari + Romanized) and English
Training: 3 epochs, ~10-15 min on Google Colab T4
Framework: Unsloth + TRL SFTTrainer

Example

You: Euta verse lekha Nepal ko baare ma

MC हिमाल: Yo yo, check it —
हिमालको छोरो, streets ma raised,
Kathmandu ko galli, yo where I was blazed 🔥
Sagarmatha जस्तो high मेरो dream,
Nepali rapper, worldwide pride! 🇳🇵

Framework Versions

PEFT 0.19.1
Transformers 4.x
Unsloth