A LoRA adapter for Qwen2.5-7B-Instruct that answers in Murad Takla style — the Bangladeshi internet meme dialect of intentionally garbled Banglish.

Trained with Unsloth in two stages (QLoRA, 4-bit base):

SFT on 60 chat examples (system/user/assistant messages).
DPO on 60 preference pairs (chosen: Murad Takla garbled Banglish, rejected: plain Banglish).

Usage

python

from unsloth import FastModel
from unsloth.chat_templates import get_chat_template

model, tokenizer = FastModel.from_pretrained(
    "ahr100007/takla-gpt",
    max_seq_length=2048,
    load_in_4bit=True,
)
tokenizer = get_chat_template(tokenizer, chat_template="qwen-2.5")
FastModel.for_inference(model)

messages = [
    {"role": "system", "content": "Tumi ekjon Murad Takla chatbot. Sob uttor Murad Takla style e dao."},
    {"role": "user", "content": "Kemon acho?"},
]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
outputs = model.generate(input_ids=inputs, max_new_tokens=512, temperature=0.8, top_p=0.95)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

The model expects the system prompt it was trained with: Tumi ekjon Murad Takla chatbot. Sob uttor Murad Takla style e dao.

Training details

Base: unsloth/Qwen2.5-7B-Instruct (loaded as 4-bit bnb)
LoRA rank 16, alpha 16, no dropout
SFT: 5 epochs, lr 2e-4, cosine schedule, loss on assistant turns only
DPO: 3 epochs, lr 5e-6, beta 0.1