BanglaNews (LoRA)

Links

• Demo (Space): https://huggingface.co/spaces/sakibalfahim/BanglaNews
• Code (GitHub): https://github.com/sakibalfahim/BanglaNews
• This model: https://huggingface.co/sakibalfahim/BanglaNews

Tasks

1. Write article: category + headline -> news body
2. Make headline: news body -> headline

Limitations (important)

• Hobby / short Kaggle T4 run (~300 + ~200 LoRA steps), not full multi-epoch training.
• Partial data coverage; generation length was capped in eval/demo.
• Automatic metrics are modest; outputs can be short.
• Not a production newsroom system. Further long fine-tuning was not pursued in this phase.
• ZeroGPU demo: queue, cold start, daily free GPU quota.

Load

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

bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.float16)
base = AutoModelForCausalLM.from_pretrained(
    'unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit',
    device_map='auto', quantization_config=bnb)
tok = AutoTokenizer.from_pretrained('sakibalfahim/BanglaNews')
model = PeftModel.from_pretrained(base, 'sakibalfahim/BanglaNews')

Data

Training summary

• Unsloth QLoRA, r=16, alpha=32, max_seq_length=2048, packing
• Adapters only (~160MB); not a full 8B merge
• Code/docs/metrics text: https://github.com/sakibalfahim/BanglaNews