README: rename _v3 section to Round 3 for consistency

README.md

@@ -73,7 +73,8 @@ run_01  Transformer + BPE 16K + FP16          ← absolute baseline
   │                                         └─ run_09  + PFNet (pfnet_hidden=32)
   │                                               └─ run_10  + per-channel decay
   │
-  └─ r2 re-runs at 15 tok/param: run_04_r2, run_07_r2
+  ├─ r2 re-runs at 15 tok/param: run_04_r2, run_07_r2
+  └─ r3 step-by-step checkpoints (v3 run): run_XX_v3/
 ```
 
 ### Round 1 — 2 tok/param
@@ -131,11 +132,12 @@ Training strategy is the only variable across runs.
 
 ---
 
-## `_v3` high-resolution folders
+## Round 3 — step-by-step checkpoints (`_v3` folders)
 
-Dense step-by-step checkpoints from the v3 training run.
-Each file contains the full **bf16 model + bf16 optimizer state** — suitable for resuming training
-or inspecting the loss landscape at fine granularity.
+Dense step-by-step checkpoints saved during the v3 training run.
+Each file contains the full **bf16 model + bf16 optimizer state** — suitable for resuming
+training at any logged step or inspecting the loss landscape at fine granularity.
+No `_final.pt` files here: these are pure training-resume archives.
 
 | Folder | Step checkpoints | Max step | Per-file size | Total size |
 |---|---|---|---|---|
@@ -152,7 +154,7 @@ or inspecting the loss landscape at fine granularity.
 | `run_14_v3/` | 75 | 9,775 | 61.3 MB | 4597 MB |
 | `run_15_v3/` | 70 | 2,350 | 61.5 MB | 4307 MB |
 
-> **Total _v3 storage:** 159.3 GB
+> **Round 3 total storage:** 159.3 GB
 
 ---