Shakespeare Model Training Progress | Generated by AI
Working! After just 500 iterations on a 10.65M parameter model, it’s generating Shakespearean-sounding (if somewhat garbled) text. Here’s the summary:
Training results:
- Model: 10.65M params, 6-layer character-level GPT
- Loss: 4.28 → 1.71 (val) in 500 steps (~30s total)
- Checkpoint saved to
out-shakespeare-char/
Samples with prompt "To be or not to be" show the model has learned:
- English words and basic grammar
- Some Shakespearean vocabulary (“Madam”, “noble”, “thee/thou”)
- Line breaks like stage dialogue
- But still incoherent at 500 iters — train to 5000 for much better quality
To get proper Shakespeare-quality text, run the full 5000 iters:
python3 train.py config/train_shakespeare_char.py