Model Training Completion Summary | Generated by AI
Training finished! Here’s the final report:
Done ✓ – all 87,000 steps, ~30 hours (1,802 min) on RTX 4070 12GB
Final stats:
- Val bpb: 0.372 – solid for a 286M model on code
- Final loss band: ~1.06-1.13 (still bouncing)
- Peak VRAM: 6,240 MiB (only 52% of 12GB – room to grow batch size)
- Checkpoint:
~/.cache/nanochat/base_checkpoints/d12/model_087000.pt
Sample outputs (base model, no SFT):
The capital of France is the capital of France, and the capital of France is the capital of
The chemical symbol of gold is a number between 0 and 9.
The opposite of hot is a bit more complicated than it is.
If 5*x + 3 = 13, then x is 5*x + 3.
Classic base-model behavior – recognizes the pattern structure but doesn’t know facts. It’s learned Python syntax and text structure (capitalize proper nouns, equation patterns, list enumerations), but it’s a token predictor, not a Q&A bot. It needs SFT to turn this into a useful assistant.
Script printed next steps:
Evaluate: python -m scripts.base_eval --device-batch-size=8
Chat: python -m scripts.chat_cli -p 'Write a Python function to reverse a linked list'
Want to run eval or try the chat CLI?