NBA Play-by-Play Summarization with T5-Gemma

Overview

This repository builds a custom dataset of NBA play-by-play (pbp) logs and finetunes T5-Gemma small model to generate concise but engaging quarter summaries of an NBA game. The pipeline covers scraping, preprocessing, and supervised fine-tuning.

Project structure

• NBA_pbp_scraper.ipynb
  Scrapes official NBA play-by-play logs and stores actions (shots, fouls, turnovers, rebounds, etc.) in structured form.
• pbp_preprocessing.ipynb
  Cleans and formats scraped data, splits games by quarter, and pairs quarter pbp text with a ground-truth summary in a kaggle dataset
• finetuning_nba.ipynb
  Finetunes google/t5-gemma-small using Hugging Face training utilities with experiment tracking via Weights & Biases

Dataset

• Source: NBA Quarter Play Summaries (custom-built)
• Granularity: Per quarter
• Link: https://www.kaggle.com/datasets/shrishtiroy6/nba-quarter-play-summaries
• Format example:
  • Input: Time TIMBERWOLVES Score Lead Warriors 12:00 Start of Period (9:16 PM) 12:00 Possession: Timberwolves 30-23 +7 11:44 MISS N.Alexander-Walker 25' 3PT Jump Shot 11:41 Q.Post REBOUND 11:34 30-25 +5 J.ButlerIII Driving Layup 11:34 N.Reid S.FOUL (P1, T1) (S.Wright) 11:34 30-26 +4 J.ButlerIII Free Throw 1 of 1
  • Output: The first quarter was tightly contested, with the Warriors leading thanks to strong three-point shooting from Curry and Hield.

Model training

• Base model: google/t5-gemma-small (60M params)
• Uses LoRA finetuning and 4-bit quantization to compress model weights
• Typical settings used:
  • Per-device batch size: 1 (with gradient accumulation)
  • Optimizer: adamw_torch_8bit
  • Mixed precision: bf16
  • Logging: Weights & Biases
  • GPU: Tesla P100
• Development goal: Overfit small examples to validate training loop, then scale to full dataset and evaluate with ROUGE/BLEU.