# SciHigh 2026 Subtask 1 — Final Pipeline

Team Yushkk99. Task: generate research highlights from paper abstracts
(MixSub corpus, 19,785 ScienceDirect papers). Official metric: ROUGE-L F1.

## Results (validation, 1,985 rows, plain rougeL / no stemming)

| system | val ROUGE-L |
| --- | --- |
| FIRE-2025 winning system (reference) | 23.45 |
| bart-large-cnn, original data | 24.37 |
| bart-large-cnn, repaired data + titles (run 1) | 26.37 |
| **Qwen2.5-72B QLoRA, repaired data + titles (run 2)** | **28.33** |

## Pipeline

Data enrichment (all sources free and sanctioned; test rows touched only
via abstract/title metadata — never highlight fields):

1. `recover_abstracts.py` — ~40% of corpus abstracts are truncated
   mid-sentence. Each row is matched to its ScienceDirect PII (verified
   100% by abstract-text join against the dataset's source repository),
   resolved to a DOI via Elsevier's keyless Article API, and its complete
   abstract fetched from Semantic Scholar. A label-free acceptance filter
   (recovered text must contain >=90% of the truncated text's tokens)
   guarantees same-paper extensions and structurally excludes records
   containing highlight fields. Acceptance: ~95% of truncated rows.
2. `fetch_titles.py` — paper titles for 100% of rows via the same keyless
   PII lookup.
3. `build_datasets.py` — rebuilds all splits with recovered abstracts
   normalized to the corpus's punctuation-stripped format (transform
   validated token-for-token against 5,481 known prefixes, mean agreement
   0.987), then prepends titles: `<title> | <abstract>`.

Modeling:

4. `train_qwen_qlora.py` — one-epoch QLoRA SFT (4-bit NF4, LoRA r=16 on
   attention+MLP) of a Qwen2.5 Instruct base on the 15,960-pair titled
   pool, chat-formatted, greedy decoding. The identical script trains the
   7B (24GB GPU) and 72B (80GB GPU) variants; only `--base` changes.
5. `infer_test.py` — greedy test-set inference producing the submission CSV.
6. `validate_submission.py` — format gate: row count, column schema, no
   empty predictions, length statistics.

Run 1 (bart lineage) reproduces via the Kaggle kernels
(`res-3-title-prefix`) recorded in the project artifacts.

## Models

- `Yakk99/scihigh2026-subtask1-qwen72b-qlora` — run-2 adapter (72B)
- `Yakk99/scihigh2026-subtask1-qwen7b-qlora` — 7B pilot adapter
- `Yakk99/scihigh2026-subtask1-bart-titled` — run-1 full model

## Negative results (tested and closed, full logs in project artifacts)

BRIO contrastive calibration (two decode-matched attempts, ~0 delta);
MBR/DPO preference signal on the 72B pool (−2.0 — sampled candidates
cluster below the greedy mode of a well-trained model); oracle bullet
reordering (+0.0002); GRPO (rejected on published scientific-domain
negatives); RAG (measured negative on this exact corpus by a FIRE-2025
team).