feat(data): add recipe-driven language supervision

This commit is contained in:
Pepijn
2026-07-28 10:04:28 +02:00
parent 95211b98f1
commit 3f093d8927
21 changed files with 779 additions and 57 deletions
@@ -0,0 +1,13 @@
# Paper-style joint sequence (pi0.5 §IV-B): one sample supervises the subtask
# text with CE and, because the assistant turn is part of the prefix, conditions
# the FAST and flow action losses on the same annotated subtask in one forward.
# The supervised span is attended causally; the action losses see task + subtask.
#
# Pair with `--policy.joint_subtask_conditioning=true` at inference so the flow
# prefix reproduces this layout (task turn with state + causal generated subtask).
# Samples without a `subtask` annotation fall back to a plain task-prompt
# low-level sample via `if_present`.
messages:
- {role: user, content: "${task}", stream: low_level}
- {role: assistant, content: "${subtask}", stream: low_level, target: true, if_present: subtask}