File size: 641 Bytes
208bdf2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
---
license: mit
base_model: facebook/bart-large
tags:
  - summarization
  - scientific-text
  - research-highlight-generation
  - lora
language:
  - en
---

# 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.