SciHigh-2026 Subtask 1 — Qwen2.5-72B QLoRA (run 2)
QLoRA adapter (4-bit NF4 base, LoRA r=16 on attention+MLP).
Fine-tuned for the FIRE 2026 SciHigh shared task, Subtask 1: generating research highlights from scientific paper abstracts (MixSub corpus).
- Input format:
<paper title> | <abstract>(corpus punctuation-stripped style; abstracts repaired via DOI-verified Semantic Scholar recovery) - Output: 3-5 highlight sentences, ~55 words, greedy decoding
- Training: one epoch on 15,960 pairs, A100-80GB
- Validation ROUGE-L F1 (plain LCS, no stemming): 28.33
Repository contents
| file | purpose |
|---|---|
train_final.py |
standalone training recipe that produced this model |
train_final.ipynb |
notebook version of the same recipe |
infer_test.py |
test-set submission CSV generation |
final/ |
full data pipeline: abstract recovery, title fetch, dataset build, submission validation |