AstraGenX: SciHigh2026 Task 1 - Research Highlight Generation
Model Details
AstraGenX-SciHigh2026-Task1 is a fine-tuned sequence-to-sequence model designed to generate concise, highly accurate research highlights from scientific abstracts. It was developed as the official submission for Task 1 of the SciHigh2026 competition.
- Developed by: Team AstraGenX
- Model type: Transformer-based sequence-to-sequence generation
- Language(s): English
- Finetuned from model:
facebook/bart-base
Uses
Direct Use
This model is intended to take a dense scientific abstract as input and output a short, bullet-point style summary (highlights) that captures the core methodology, findings, and significance of the research.
Downstream Use
It is optimized for automated academic workflows, literature review assistance, and fast comprehension of dense STEM-related papers.
Training Details
Training Data
The model was fine-tuned on the MixSub dataset provided by the SciHigh2026 organizers, consisting of paired scientific abstracts and their corresponding research highlights.
Training Procedure
The model was trained using Hugging Face's Trainer API on a GPU-accelerated environment, utilizing the following constraints specified by the competition:
- Max Input Length: 512 tokens
- Min Generation Length: 30 tokens
- Max Generation Length: 150 tokens
- Generation Strategy: Beam Search (num_beams=4) to ensure high-quality lexical overlap and semantic coherence.
Evaluation
The model's performance is measured against the competition baseline (Pegasus ROUGE-L F1: 23.45%) using the following official automatic metrics:
- ROUGE (1, 2, L): Lexical overlap and exact match accuracy.
- METEOR: Semantic similarity incorporating synonym and stemming matching.
- BERTScore: Deep semantic equivalence using contextual embeddings.
