fix: align SFT/DPO figures and text with actual training data

Browse Source

- SFT: 1,000 steps/WSD → ~3,800 steps/cosine; loss ~2.5→1.6 → ~2.1→1.5
- DPO: fix preference-loss narrative to match training-loss curve;
  correct initial grad-norm (~50 → ~200)
- Table 2: peak LR 1.5e-4 → 2.0e-4 (matches pt_metric.png)
- Clarify scheduling: pretraining=WSD, SFT+DPO=cosine
- Add disclaimer to ckpt_weight_density_per_run caption for legacy iter labels

...

This commit is contained in:

ViperEkura

2026-07-24 06:15:42 +08:00

parent 4b37f289c0

commit a7bbc7b29f

3 changed files with 30 additions and 30 deletions

Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL

Download Patch File Download Diff File Expand all files Collapse all files

data/dpo_metric.png

BIN

Binary file not shown.

Side by Side Swipe Overlay Before Width:  |  Height:  |  Size: 92 KiB After Width:  |  Height:  |  Size: 86 KiB

data/sft_metric.png

BIN

Binary file not shown.

Side by Side Swipe Overlay Before Width:  |  Height:  |  Size: 85 KiB After Width:  |  Height:  |  Size: 84 KiB