Emender E97 4B exact training checkpoints
Private operational repository for durable, checksummed training checkpoints of the 4,045,972,080-parameter E97 recurrent base campaign.
These are raw torch.save training checkpoints. They contain model and
Schedule-Free optimizer state and use Python pickle. Load only from this trusted
repository and verify the adjacent SHA256SUMS first. They are not portable
model releases.
LATEST.json identifies the newest published durable. Each checkpoint directory
contains immutable metadata, original training arguments, and checksums.
Durable checkpoints
| Step | Accepted tokens | World | Sampler | Last-100 loss | Status |
|---|---|---|---|---|---|
| 12,800 | 6,710,886,400 | 8 | counter-v1 | 2.8143 | Lambda seed |
| 24,448 | 99,723,771,904 | 2,048 | counter-v2 | 2.3981 | Frontier final authority |
The final Frontier authority uses 256 node-local eight-rank DDP islands, B4 per
rank (effective B32 per island), and K32 DiLoCo. Payload 5384001 and collector
5384002 completed 0:0; atomic publication and independent mmap reload passed.
Its SHA-256 is
3ace004251643acf2e7c7f720e8f29968ad0a483441553c0c885b87b3df84568.
Exact sampler continuation is fixed to each checkpoint's recorded data world, stream origin, corpus, tokenizer, and transition chain. Cross-world resume requires an explicit counter-v2 phase transition; silent relabeling is invalid.