Text Generation Model with GPT-2

General-purpose language models drift off-topic and hallucinate specifics when asked to write within a technical domain. The project asked whether a modest fine-tune on curated data could measurably reduce that drift.

• Assembled and cleaned a domain corpus of roughly 40k passages, deduplicating and normalising formatting.
• Fine-tuned GPT-2 in PyTorch with staged learning-rate schedules and early stopping on validation perplexity.
• Built a human evaluation rubric scoring factuality, coherence and tone against baseline outputs.
• Wrapped the model in a small API layer so reviewers could compare generations side by side.

• Validation perplexity reduced by 31% against the base checkpoint.
• Human reviewers preferred fine-tuned output in 7 of 10 blind comparisons.
• Written up as the graded senior project and released as an open-source repository.