nit_patna_2026 — SciHigh 2026 Task 1 (Research Highlight Generation)
Two-stage pipeline: select, then rewrite.
Stage 1 — sentence selection (selector/). SciBERT (allenai/scibert_scivocab_uncased)
fine-tuned as a regressor: input "{i}/{n} sentence" [SEP] abstract, target = max ROUGE-L F1
between the sentence and the reference highlight bullets. At inference, sentence relevance
from this model is combined with a sentence-sentence cosine-similarity matrix
(all-mpnet-base-v2 embeddings) by Maximal Marginal Relevance:
argmax lam*rel(s) - (1-lam)*max_t sim(s,t), lam = 0.8, k = 4 sentences.
Stage 2 — rewriting (generator/). QLoRA adapter (r=16, 4-bit NF4 base) on
Qwen/Qwen3-4B-Instruct-2507, trained on 4000 MixSub training papers to turn the selected
sentences into 3-5 highlights. Training selections were cross-fitted (each half of train
selected by a scorer trained on the other half). Greedy decoding, max 128 new tokens, repeated 6-grams within the output blocked, near-duplicate bullets dropped, max 5 bullets.
Prompt:
Below are the key sentences selected from the abstract of a scientific paper.
Write the paper's research highlights: 3 to 5 short sentences, one per line, each under 85 characters.
Key sentences:
{src}
Team: nit_patna_2026 (NIT Patna). Data: MixSub-SciHigh (FIRE 2026).