Commit
Single-term loss diagnostics presets
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
modified
scripts/train.py
+5 −0
@@ -30,6 +30,11 @@ PRESETS = { | ||
| 30 | 30 | "m3_noadv": dict(arch="m3", loss=dict(style=0.3, con=0.2, sem=0.1, orth=0.1, fact=0.5)), |
| 31 | 31 | "m3_nofact": dict(arch="m3", loss=dict(style=0.3, adv=0.3, con=0.2, sem=0.1, orth=0.1)), |
| 32 | 32 | "m3_nocon": dict(arch="m3", loss=dict(style=0.3, adv=0.3, orth=0.1, fact=0.5)), |
| 33 | + # single-term diagnostics (pilot: every M3 variant with auxiliary losses generalised far worse than token-only) | |
| 34 | + "m1_fact": dict(arch="m1", loss=dict(fact=0.5)), | |
| 35 | + "m3_factonly": dict(arch="m3", loss=dict(fact=0.5)), | |
| 36 | + "m3_cononly": dict(arch="m3", loss=dict(con=0.2)), | |
| 37 | + "m3_semonly": dict(arch="m3", loss=dict(sem=0.1)), | |
| 33 | 38 | } |
| 34 | 39 | |
| 35 | 40 | |
| 36 | 41 | |