fix(finetune): make the inert LIBERO drop-gate honest + fail-closeable#231
Open
rylinjames wants to merge 1 commit into
Open
fix(finetune): make the inert LIBERO drop-gate honest + fail-closeable#231rylinjames wants to merge 1 commit into
rylinjames wants to merge 1 commit into
Conversation
Audit §3.4 / Part 1 #10. libero_drop_gate runs teacher-vs-student LIBERO rollouts and aborts the ship if task-success drops too far. But the rollout harness (tether.libero_harness) isn't shipped (LIBERO was archived 2026-04-17), so _run_teacher_student_rollouts ALWAYS raises _LiberoUnavailable and the gate silently skips — every distill ships with status "ok" having verified nothing, and a SKIP is indistinguishable from a PASS. A gate that can't fail is worse than no gate. Can't ship the sim harness here, so make the gap honest and give callers a fail-closed switch instead: - PostprocessReport gains libero_gate_status; the hook records the outcome on EVERY path (passed / failed / crashed / skipped_unavailable / skipped_disabled / skipped_phase / skipped_missing_inputs). A skip is now visible and auditable, never read as a pass. - New extra_lerobot_args.libero_gate_require (default False): when True, an unavailable harness ABORTS the ship rather than silently skipping — so a production distill can demand real task-success verification. - Docstring states plainly that the harness is absent in OSS builds and how to fail closed. Tests (tests/test_libero_drop_gate.py, 6): unavailable→skipped_unavailable + ships by default; require=True→force_abort; pass/fail record status + abort; skip + non-distill phase recorded. ruff clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Audit §3.4 / Part 1 #10.
Problem
libero_drop_gateis supposed to run teacher-vs-student LIBERO rollouts and abort the ship if task-success drops too far. But the rollout harness (tether.libero_harness) isn't shipped (LIBERO was archived 2026-04-17), so_run_teacher_student_rolloutsalways raises_LiberoUnavailableand the gate silently skips. Every distill ships with status "ok" having verified nothing — and a SKIP is indistinguishable from a PASS. A gate that can't fail is worse than no gate.Fix
The sim harness can't be shipped here, so make the gap honest and give callers a fail-closed switch:
PostprocessReportgainslibero_gate_status; the hook records the outcome on every path (passed/failed/crashed/skipped_unavailable/skipped_disabled/skipped_phase/skipped_missing_inputs). A skip is now visible and auditable — never read as a pass.extra_lerobot_args.libero_gate_require(defaultFalse): whenTrue, an unavailable harness aborts the ship rather than silently skipping — a production distill can demand real task-success verification.Tests
tests/test_libero_drop_gate.py(6): unavailable →skipped_unavailable+ ships by default;require=True→force_abort; pass/fail record status + abort; skip + non-distill phase recorded. ruff clean.🤖 Generated with Claude Code