feat(spaghetti): per-trajectory KL/p95 normalization + shared y-axis

Browse Source

- Add --normalize-kl / --calib-tokens (default on, 20 tokens) to
  spaghetti_kl_alive.py: each trajectory is divided by its own
  p95(KL[:calib_tokens]). Calibration target collapses to y=1.
- Share y-axis across all alpha panels (global p99 ymax) for
  direct cross-alpha comparison.
- Add OLMo-2 1B w=4096 figure to figs/ and a README section
  documenting the result, including the unexplained 'random
  dead traces at low alpha' observation.

...

This commit is contained in:

wassname

2026-05-08 08:17:21 +08:00

parent 7d3fd37743

commit dbff43cffb

3 changed files with 105 additions and 16 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

figs/spaghetti_olmo2_1b_w4096_norm.png

BIN

Binary file not shown.

Side by Side After Width:  |  Height:  |  Size: 401 KiB