# 2026-06-06 - Release note for `step_11000` from the short-fast-decay 11k web/wiki run

## Release candidate

- run: `20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki`
- config: `configs/testing/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki.yaml`
- chosen checkpoint: `/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_11000.pt`
- intended HF repo: `nazdef/gpt2small-en-it-nanochat-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki-step11000`

## Why this checkpoint

This run finished cleanly at `11000/11000`, and the final checkpoint is both:

- the best saved online validation checkpoint:
  - `validation_loss=3.8305816711`
  - `validation_perplexity=46.0893392775`
- the best checkpoint under the repo-native CPU benchmark across `7000/8000/9000/10000/11000`

Primary benchmark ranking by `val_loss_mixed` lower-is-better:

1. `step_11000`
   - `mixed=5.2270`
   - `en=4.9931`
   - `it=4.0279`
2. `step_7000`
   - `mixed=5.2277`
   - `en=5.0606`
   - `it=4.0280`
3. `step_10000`
   - `mixed=5.2677`
4. `step_8000`
   - `mixed=5.3010`
5. `step_9000`
   - `mixed=5.3963`

The win over `step_7000` is real but tiny:

- `delta mixed (11000 - 7000) = -0.0006882`

Supporting signals still lean toward `11000` as the preserve/publish point for this run:

- `loop_rate`: `0.40` vs `0.60` at `7000`
- `distinct_2`: `0.4990` vs `0.4338` at `7000`
- `ppl_en`: `147.39` vs `157.69` at `7000`
- `ppl_it`: effectively tied (`56.1452` vs `56.1479`)

## Cross-run reading

Against the comparable public champion checkpoint tracked from this workspace:

- champion: `20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki step_7000`
- champion `val_loss_mixed=5.2158`
- this release `val_loss_mixed=5.2270`
- delta vs champion: `+0.0112`

Operational reading:

- this `step_11000` is the best checkpoint of the short-fast-decay run
- it is also the most stable-looking checkpoint among the recent short-fast/new candidates in this family
- it still does **not** take the overall crown from `earlydecay7000 step_7000`; it misses by a small but real margin on the primary benchmark metric

## Caveats

- this is a comparative/operational win, not a claim of polished free-form generation quality
- EN improves more clearly than IT
- generations remain repetitive enough that "best recent candidate" and "best overall model" are not the same claim