from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-4B")
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-4B", dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, "sinhala-nlp/Qwen3.5-4B-NSINA-Headlines-si")

Prompts use the base model's chat template with thinking disabled, and the assistant response begins with the Headline: prefix. Articles are trimmed to 2500 characters and then budgeted to fit max_seq_len from the lead.

Training


Training articles	8000
Instruction language	`si`
Epochs	1.0
Effective batch size	16
Learning rate	0.0002
Max sequence length	2560
LoRA r / alpha / dropout	16 / 32 / 0.05
Thinking during training	False

Evaluation

First 1000 instances of the NSINA-Headlines test split, ROUGE F1 x100 with whitespace tokenization (the default rouge_score tokenizer strips non-ASCII and zeroes out every Sinhala score):

Metric	Score
ROUGE-1	28.22
ROUGE-2	13.14
ROUGE-L	27.48

Licence

This adapter inherits the licence of the base model; check the base model card before redistributing. The training data is NSINA-Headlines, derived from scraped Sri Lankan news content -- verify its terms on the dataset card, as they may be more restrictive than the base model's licence.