# 2026-06-03 - Cross-run leaderboard update for the published `step_7000` release

This note records the later five-way cross-run re-evaluation that was completed after the original `2026-06-01` release.

## What changed

The repo-native CPU benchmark suite `configs/eval/20260521_pretrain_minimal_en_it_webwiki_step11000.yaml` was rerun across five previously selected winner checkpoints from distinct GPT-2-small EN/IT web/wiki run families.

Primary decision metric:

- `val_loss_mixed` lower is better

Final leaderboard:

1. `earlydecay7000 step_7000`
   - `val_loss_mixed=5.2158`
2. `earlydecay3500 step_7000`
   - `val_loss_mixed=5.2358`
3. `lr2e4 cosine step_7000`
   - `val_loss_mixed=5.3558`
4. `lr2e4 WSD step_11000`
   - `val_loss_mixed=5.3576`
5. `lr1e4 cosine step_14000`
   - `val_loss_mixed=5.4493`

## Reading

The winner stayed unchanged. That means the published `earlydecay7000 step_7000` release remains the current best benchmark checkpoint across the comparable public GPT-2-small EN/IT web/wiki slice tracked from this workspace.

This should still be read narrowly:

- it is the best current checkpoint under the comparable benchmark protocol
- it is not a claim that free-form generations are already strong
- qualitative outputs remain weak and repetitive across the leaderboard family

## Added supporting artifacts

- `cross_run_leaderboard_report_20260603.md`
- `cross_run_leaderboard_comparison_20260603.json`
- `cross_run_leaderboard_summary_20260603.json`