README.md

metadata

license: mit
tags:
  - nanochat
  - training-dynamics
  - checkpoints

nanochat d24 training checkpoints (2026-08-29)

Full intermediate checkpoint trajectories for three nanochat d24 pretraining runs.

folder	run	steps saved	model_*.pt
d24/	base	every 50 steps, up to 5568	112
d24_6shard_10ep/	6-shard, 10 epochs	every 26 steps, up to 2590	100
d24_13shard_10ep/	13-shard, 10 epochs	every 56 steps, up to 5600	100

Each folder contains:

• model_XXXXXX.pt — model weights at that step (~3.9 GB each)
• optim_XXXXXX_rank{0..7}.pt — optimizer state at the final step (resume training)
• meta_XXXXXX.json — per-checkpoint metadata
• val_metrics.jsonl — validation curve for the run