How I Fine-Tuned GPT-2 for Cricket Commentary

Why this project mattered

Sports commentary is emotional, fast, and contextual. Generic language models sound flat when they describe cricket moments.

I wanted a model that could generate short, punchy lines that feel like live commentary while still being grounded in what happened in the match.

Dataset and preprocessing

I collected a large commentary dataset and normalized inconsistent text patterns such as over formats, team names, punctuation, and special symbols.

I removed noisy rows, duplicated snippets, and low-value boilerplate so the model would focus on useful commentary style instead of artifacts.

• Built a cleaned dataset focused on meaningful ball-by-ball context
• Standardized abbreviations and match metadata
• Created train and validation splits to avoid leakage

Modeling approach

I used GPT-2 fine-tuning for generation quality and compared multiple prompt templates to stabilize tone.

I also tested sequence length and decoding settings because cricket commentary quality drops quickly when outputs become repetitive.

• Prompt format had a major impact on output style
• Sampling settings changed creativity versus factual consistency
• Short output targets gave the best production-like results

Training and evaluation

Training quality improved when I balanced learning rate, context window, and batch settings for the GPU budget available.

I evaluated with both offline metrics and manual review, since sports language needs human judgment for excitement and readability.

• Tracked loss curves across checkpoints
• Compared outputs on the same held-out match contexts
• Selected the checkpoint with best readability and style match

Key outcomes

The final model produced sharper, more idiomatic commentary lines than baseline prompting on a general model.

The biggest lesson was that dataset quality and prompt structure mattered more than trying complex model changes too early.