Fine-Tuned BART-Large for SciHigh 2026 (Task 2)

This repository contains a fine-tuned BART-Large model trained specifically for the SciHigh 2026 (Subtask 2) competition to automatically generate concise and accurate scientific titles given research paper abstracts.

Model Details

• Developed by: Vino1502
• Model Type: Sequence-to-Sequence (Seq2Seq) Transformer
• Language: English
• Base Model: facebook/bart-large
• Task: Scientific Abstract-to-Title Generation (SciHigh 2026 - Subtask 2)

Uses

Direct Use

This model is intended for scientific title generation. Given a research paper abstract as input, the model generates a concise scientific title summarizing the abstract's core findings.

How to Get Started with the Model

You can load and run inference directly with this model using the following code:

import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

# Configuration
MODEL_REPO_ID = "Vino1502/scihigh-2026-task2-bart"

# Load Tokenizer & Model directly from HF Hub
tokenizer = AutoTokenizer.from_pretrained(MODEL_REPO_ID)
model = AutoModelForSeq2SeqLM.from_pretrained(
    MODEL_REPO_ID,
    dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    device_map="auto"
)
model.eval()

# Sample Inference
abstract_text = "Your scientific abstract goes here..."

# Tokenize and place tensors on the model's device
inputs = tokenizer(
    abstract_text, 
    return_tensors="pt", 
    max_length=512, 
    truncation=True
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_length=64,
        num_beams=2,
        early_stopping=True
    )

predicted_title = tokenizer.decode(outputs[0], skip_special_tokens=True)
print("Predicted Title:", predicted_title)