fix(pi052): default flow loss weight to ten

This commit is contained in:
Pepijn
2026-07-15 15:05:13 +02:00
parent 0a7b21cdd0
commit 6a89c7be45
2 changed files with 3 additions and 6 deletions
+1 -1
View File
@@ -111,7 +111,7 @@ the expected prompt, text target, and action endpoints before scaling up.
| -------------------------------- | -------------------------: | ------------------------------------------------------------------ |
| `policy.recipe_path` | `recipes/subtask_mem.yaml` | Selects the language/action objective mixture |
| `policy.text_loss_weight` | `1.0` | Language-head cross-entropy weight; `0` disables text training |
| `policy.flow_loss_weight` | `5.0` | Continuous action flow-loss weight |
| `policy.flow_loss_weight` | `10.0` | Continuous action flow-loss weight |
| `policy.enable_fast_action_loss` | `true` | Adds discrete FAST action-token supervision |
| `policy.fast_action_loss_weight` | `1.0` | FAST cross-entropy weight |
| `policy.knowledge_insulation` | `true` | Blocks action-loss gradients through the VLM K/V path |
@@ -48,11 +48,8 @@ class PI052Config(PI05Config):
"""Weight on the LM-head cross-entropy term. Set to ``0`` to disable
text training entirely (reverts to flow-only / π0.5 behaviour)."""
flow_loss_weight: float = 5.0
"""Weight on the action-expert flow-matching term. ``5.0`` — a milder
flow:text split than the paper's α=10, since the flow-only
``low_level`` recipe already gives the action expert frequent
gradient. Lower it further if the LM head still underfits."""
flow_loss_weight: float = 10.0
"""Weight on action-expert flow matching relative to text supervision."""
# Backbone training ---------------------------------------------------
unfreeze_lm_head: bool = True