Confirm v3 rejection with extended fixed-MATH alpha line search
Browse files
candidates/budgie-alignment-v2/reasoning-sft-then-rl-v3/recovery/V3_FINAL_REPORT.md
CHANGED
|
@@ -28,3 +28,7 @@
|
|
| 28 |
| FOLIO fixed | 13/30 | 13/30 |
|
| 29 |
|
| 30 |
The SFT direction gives a small information gain, but does not improve random math and regresses fixed GSM/MATH. Smaller alphas do not improve held-out information management. Under the predeclared protocol, RL may run only after an accepted SFT anchor, so follow-on RL was correctly skipped.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
| FOLIO fixed | 13/30 | 13/30 |
|
| 29 |
|
| 30 |
The SFT direction gives a small information gain, but does not improve random math and regresses fixed GSM/MATH. Smaller alphas do not improve held-out information management. Under the predeclared protocol, RL may run only after an accepted SFT anchor, so follow-on RL was correctly skipped.
|
| 31 |
+
|
| 32 |
+
## Extended MATH line search
|
| 33 |
+
|
| 34 |
+
After the initial decision, alphas .025, .05, .1, .2, .5, and 1.0 were also evaluated on the fixed MATH rows. Their scores were 1, 1, 2, 2, 2, and 2 out of 15; alpha .005 scored 2/15 versus the leader 3/15. No point on the tested task-vector line preserves fixed MATH, confirming the Stage A rejection.
|