Sentence Completion (Tiny GPT)

• Word-level tokenizer • Trained on startup • Dataset prefix match (≥2 words) • Greedy fallback