Correct the report's checkpoint-selection section: only global_step_25 exists
Browse files- training_logs/report.md +14 -0
training_logs/report.md
CHANGED
|
@@ -1,3 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
# SkyRL Standard-GRPO Training Metrics Analysis
|
| 2 |
|
| 3 |
Generated from 1 log file(s) (`--format standard`)
|
|
|
|
| 1 |
+
> **Correction — ignore the "Best Checkpoint" section below.**
|
| 2 |
+
>
|
| 3 |
+
> That section was generated without any export listing available to the parser (it reports
|
| 4 |
+
> `available exports: []`). With nothing to intersect against, its "saved-aligned" filter degraded to
|
| 5 |
+
> ranking every metric row by trailing-5 EMA, so it names step 60 and an export path
|
| 6 |
+
> `exports/global_step_60/policy/` that **does not exist**.
|
| 7 |
+
>
|
| 8 |
+
> `global_step_25` is the only checkpoint and the only export ever written for this run — verified
|
| 9 |
+
> against the object store, where `checkpoints/` and `exports/` each contain exactly one entry. The
|
| 10 |
+
> published weights are therefore the only candidate that existed, not a choice made over better ones.
|
| 11 |
+
>
|
| 12 |
+
> Note also that this job is a resumed segment: its metrics begin at step 22, so steps 1-21 belong to
|
| 13 |
+
> the predecessor run in the lineage.
|
| 14 |
+
|
| 15 |
# SkyRL Standard-GRPO Training Metrics Analysis
|
| 16 |
|
| 17 |
Generated from 1 log file(s) (`--format standard`)
|