Siddharth Sen

Runners Up(2nd place) in BrainDead ML Hackathon - Mar 2025 I contributed by training and fine-tuning the BART-Large-CNN model on metadata from over 15,000 research papers from the ArXiv dataset. I contributed by increasing the ROUGE-1, ROUGE-2 and BLEU scores of the model when compared to some benchmark models like : SummRuNNer, Pointer-Generator, Discourse Aware, BERT, T5, BART, PEGASUS, GPT-4, Longformer, and LED. Here is a description of our solution : 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗔𝗿𝘁𝗶𝗰𝗹𝗲 𝗦𝘂𝗺𝗺𝗮𝗿𝗶𝘇𝗮𝘁𝗶𝗼𝗻 – The… Show more I contributed by training and fine-tuning the BART-Large-CNN model on metadata from over 15,000 research papers from the ArXiv dataset. I contributed by increasing the ROUGE-1, ROUGE-2 and BLEU scores of the model when compared to some benchmark models like : SummRuNNer, Pointer-Generator, Discourse Aware, BERT, T5, BART, PEGASUS, GPT-4, Longformer, and LED. Here is a description of our solution : 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗔𝗿𝘁𝗶𝗰𝗹𝗲 𝗦𝘂𝗺𝗺𝗮𝗿𝗶𝘇𝗮𝘁𝗶𝗼𝗻 – The 𝙃𝙮𝙗𝙧𝙞𝙙𝙎𝙪𝙢𝙢𝙖𝙧𝙞𝙯𝙚𝙧 :We built 𝗛𝘆𝗯𝗿𝗶𝗱𝗦𝘂𝗺𝗺𝗮𝗿𝗶𝘇𝗲𝗿, an AI-powered system trained on ArXiv data, to generate concise, high-quality research summaries while preserving key insights.Built on BART, it combines a bidirectional encoder for deep contextual understanding and an autoregressive decoder for fluent summarization. Fine-tuned on structured research metadata and abstracts, it processes inputs up to 512 tokens with optimized Seq2Seq training for accuracy.• Trained & Validated on: ArXiv dataset• Fine-tuned with: Tokenization, FP16 precision, optimized learning rates Show less