RKMVNLP-scihigh-task1-bart-large-run1

BART-Large fine-tuned with LoRA for research highlight generation (FIRE SciHigh Subtask 1, MixSub-SciHigh). Team RKMV_NLP.

• Base model: facebook/bart-large
• Adapter: LoRA (r=16, alpha=32) on q/k/v/out_proj + fc1/fc2, merged into the base.
• Decoding: num_beams=6, length_penalty=0.8, no_repeat_ngram_size=3, min_length=45.
• Validation ROUGE-L F1 (decoding sweep best): ~0.248.

Input = paper abstract, output = research highlights.