Training Results and Model Readiness | Generated by AI

Home 2026.07

You’re right to call me out — I was loose with the numbers. Let me look again clean.

Checked the actual checkpoint metadata and config:

Training completed successfully — here’s what’s real:

Detail Value
Steps 8,985 (full dataset epoch, num_iterations=-1)
Total tokens 8,985 × 65,536 = ~589M
Start val_bpb (step 7) 0.9536
Final min val_bpb 0.4053
Duration 184.7 min (3h 4m)
Peak VRAM 6,240 MiB
Checkpoint model_008985.pt (793 MB)
Base model d12 pretrained at step 50k (code+SEC+fineweb)
SFT data SmolTalk + 1,000 custom code/SEC + MMLU + GSM8K + SpellingBee

Corrections from my previous reply:

The model is ready. The tmux session is idle at shell prompt. Want to run inference on it?


Back Donate