# Checkpoints

Updated 2026-10-09. Only the d64 lineage tips remain.

## Keep

| File | Step | Role |
|---|---|---|
| `bi_d64_ansbank/atom_native_step_12100000_bi_d64_ansbank.pt` | 12,100,000 | Training tip. CUDA, B=1920, compile, BPTT 256, role-loss, answer bank. Uploaded 2026-10-02. |
| `bi_d64_fds/atom_native_step_10700000_bi_d64_fds.pt` | 10,700,000 | Parent. Fractus-dataset dialogue pass. Resume source of the 12.1M run. |
| `bi_d64_day/atom_native_step_9500000_bi_d64_day.pt` | 9,500,000 | Parent of the 10.7M file. |

Resume the 12.1M file with the same flags saved in it: d=64, n_modes=64, max span 1, leak 0.05 frozen, batch 1920, field BPTT 256, role loss on. Do not restart from token 0.

## Not speech

Last published greedy samples are from 10.6M (`C'est de la fait…`, `Je veux vie trais ?`, `Alle se`). The 12.1M file has not been free-run probed in this card.

## Removed

`RELEASE/`, `gpu_*` pod copies, `bi_d64_role`, `bi_d64_rolefds`, `bi_d64_pdep`, `bi_d64_leak05`, d32 `byte_tick_bptt_*`, smoke, and intermediate ansbank steps 11.0M, 11.4M, 11.7M.

The previous card called `byte_tick_bptt_k128_mix` step 1,720,000 the chat tip. That line is deleted. It is not the current brain.