Swap the annotation VLM from Qwen3.6-35B-A3B (sparse MoE, ~3B active)
to Qwen3.6-27B (dense, 27B all-active). Per Scale's dense-captioning
study, model capacity is the #1 lever and the dominant failure is
visual grounding — both helped by ~9x more active params. Qwen3.6-27B
is a vision-language model (vision encoder, image + video), same family
so the chat template / video handling / enable_thinking=false flag are
unchanged, and at 27B dense it still fits one H200 per server, so the
two-parallel-server layout (TP=1, one per GPU) is preserved — no
throughput-layout change, just a much stronger model.
Kept: parallel_servers=2, num_gpus=2, max-model-len 32768 (the 32-frame
embedded budget is ~10k tokens, well under), gpu-mem 0.8.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Switching the plan module to embedded frames (use_video_url=false)
exposed a context overflow: at frames_per_second=2.0 with the old
max_video_frames=128 default, a 480x640 episode embeds ~128 frames ≈
33-39k vision tokens, over the model's 32768 context — every plan call
died with 'Input length exceeds maximum context length' (HTTP 400),
crashing the whole annotation job.
The video_url path never hit this because the server downsampled; the
embedded path sends every sampled frame, so the frame count is a hard
token budget.
Fix:
* config default max_video_frames 128 -> 32 (~8-10k vision tokens,
comfortable headroom for the prompt + describe/verify passes).
Frames are still sampled UNIFORMLY across the whole episode, so
longer episodes are subsampled, not truncated — full temporal
coverage preserved, just coarser density.
* run_hf_job.py: frames_per_second 2.0 -> 1.0, explicit
--plan.max_video_frames=32, with a comment explaining the token
budget and the 'do not raise toward 128 with embedded frames' rule.
Only the plan module embeds the full episode; VQA (1 frame/tick) and
interjections (4-frame window) were never at risk.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Flip PlanConfig.subtask_describe_first and subtask_verify defaults
False -> True. Every subtask annotation now runs the 3-call grounding
+ pruning chain by default, since the single-call path reliably
hallucinates steps from the task text. Costs 2 extra VLM calls/episode;
disable with --plan.subtask_describe_first=false / --plan.subtask_
verify=false on easy datasets where fewer calls matter more than
label fidelity.
run_hf_job.py: drop the now-redundant explicit flags, leave a note that
the chain is default-on and how to opt out.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The single-call 'watch video -> emit subtask JSON' pattern makes the
VLM commit to structured output before reasoning about what it saw, so
it pattern-matches the task text and hallucinates steps. Split it into
an opt-in multi-call chain that grounds first and prunes last.
New PlanConfig flags (both default False -> single-call unchanged):
* subtask_describe_first: a grounding pass narrates ONLY what is
visible in the video (no subtask JSON yet). That description is
injected into the segmentation prompt via a new {observation_block}
placeholder, so the model segments its own grounded observations
instead of the instruction text. +1 VLM call/episode.
* subtask_verify: after segmentation, an adversarial pass re-watches
the video and drops any candidate subtask it cannot see. Can only
PRUNE (never add/rewrite/move) and fails open (keeps un-verified
spans if the call returns nothing). +1 VLM call/episode.
Implementation:
* _generate_subtasks now orchestrates describe -> segment -> verify.
* Factored span cleaning into _clean_spans (shared by segment + verify
outputs); added _describe_episode and _verify_subtasks helpers.
* New prompts module_1_subtask_describe.txt (returns {description})
and module_1_subtask_verify.txt (returns pruned {subtasks}).
* module_1_subtasks.txt gains a {observation_block} slot at the top.
run_hf_job.py enables both for the RoboCasa run (3 VLM calls/episode
for subtasks). Combined with single-camera grounding + the embedded-
frame path, this is the high-quality configuration.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two fixes for 'subtasks describe actions not in the video' plus a way
to focus the whole pipeline on one camera.
ANTI-HALLUCINATION
1. _episode_video_block: when use_video_url is set but clip extraction
fails, FALL BACK to embedded frames instead of returning an empty
block. An empty block left the VLM with zero visual grounding, so
it invented subtasks from the task text alone — the likely root
cause of hallucinated steps. Now logs a warning and embeds frames.
2. module_1_subtasks.txt gains a GROUNDING preamble (overrides all
other rules): label only motion visible in specific frames; never
invent/anticipate/pad; max_steps is a CEILING not a target; atomic
demos may be exactly ONE subtask; the VIDEO is ground truth, not
the instruction text.
SINGLE-CAMERA GROUNDING
* New VqaConfig.restrict_to_default_camera (default False). When True,
the VQA module grounds on only the --vlm.camera_key stream instead
of iterating every camera — matching the plan / interjection
modules, which already use that single camera. Now the whole
pipeline can focus on one view (e.g. observation.images.base).
run_hf_job.py updated:
* use_video_url=false + frames_per_second=2.0 — embed frames directly
(most reliable; no silent text-only failure mode) with dense
grounding.
* vqa.restrict_to_default_camera=true — VQA on the single camera too.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Drops the replace_subtask_text option and the
_render_action_record_to_subtask_text renderer. Action records are now
strictly additive: when action_records.enabled=True the module emits
style='action_record' rows (the typed {verb,object,arm,grasp,dest,
mistake} schema) and NEVER rewrites the subtask text the policy
conditions on.
The render-back-to-text path was the source of corrupted subtasks
(navigation tasks produced 'move stove to stove', manipulation tasks
got spurious 'with left arm using pinch grip' suffixes). Reconstructing
natural-language subtasks from hallucinated structured fields is
inherently fragile, so the capability is removed rather than guarded.
Removed:
* ActionRecordsConfig.replace_subtask_text field
* PlanSubtasksMemoryModule._render_action_record_to_subtask_text
* the span['text'] = canonical_text overwrite in run_episode
Updated docstrings + run_hf_job.py comment accordingly. emit_record_row
(default True) is now the feature's only output.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three compounding bugs made RoboCasa annotation produce off-task
subtasks ('move stove to stove with left arm') and drifting
augmentations ('wander around the kitchen' for 'Navigate to the stove').
1. action_records.replace_subtask_text now defaults False.
Overwriting the VLM's subtask text with a reconstruction of
hallucinated {verb,object,arm,grasp,dest} fields is high-risk:
navigation / non-manipulation tasks don't fit the schema and render
to nonsense. Records are now additive by default (emit_record_row),
never silently replacing subtask text. Flip replace_subtask_text on
only for manipulation datasets verified to render cleanly.
2. _render_action_record_to_subtask_text drops a degenerate
destination that just echoes the object (verb=move object=stove
destination=stove -> 'move stove' instead of 'move stove to stove').
Also routes 'navigate' through the 'to <dest>' preposition family.
3. module_1_task_aug_axes.txt hardened: variants MUST preserve the
goal/destination. Explicitly forbids 'Navigate to the stove' ->
'wander around the kitchen'. Only wording / arm / orientation /
grasp may vary; verb meaning, object, and destination are fixed.
examples/annotations/run_hf_job.py — corrected for RoboCasa:
* derive_task_from_video=off (was =always). The dataset task string
is authoritative and is what eval conditions on; =always threw it
away, re-derived a hallucinated task from the video, and poisoned
every downstream subtask/plan row. THIS was the dominant cause.
* n_task_rephrasings=0 + task_aug_axes left off — RoboCasa eval uses
exact task strings, so augmentation is unused/harmful.
* action_records left off — manipulation schema doesn't fit atomic /
navigation tasks.
* plan_max_steps=6 to keep atomic-task decomposition tight.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
VideoFrameProvider derived its default camera and camera list from
meta.camera_keys, which mixes image- and video-stored cameras. The
clip/decode paths read videos/<key>/from_timestamp, which only exists
for video keys, so an image-stored camera sorted first (e.g.
observation.images.wrist) crashed the plan phase with a KeyError.
Restrict the list and default to meta.video_keys. Add a regression test
and point the example job at the dataset's actual video camera. Skip
bandit B607 (ffmpeg/git are intentionally resolved via PATH).
Co-authored-by: Cursor <cursoragent@cursor.com>
Heterogeneous datasets (different tasks/scenes across episodes) don't
share a single small subtask + memory vocabulary, so the canonical
vocabulary phase narrowed every episode to the wrong target distribution.
Flip the example to free-form generation by default and document the
``--vocabulary.enabled=true`` switch for homogeneous datasets where the
canonical vocabulary still helps the downstream policy.
No pipeline-code changes: ``VocabularyConfig.enabled`` already gates
phase 0 (see ``executor.py:_run_vocabulary_phase`` and
``VocabularyConfig`` docstring) and falls back to free-form generation.
Co-authored-by: Cursor <cursoragent@cursor.com>
Hardcoding ``n_subtask_target=10`` and ``n_memory_target=6`` baked task
complexity into the config — a simple pick-and-place needs ~6, a
multi-step recipe needs ~20. The VLM already sees the clips, so let it
pick the count itself from what's recurring across episodes.
Drop both knobs from ``VocabularyConfig`` and the ``module_0_vocabulary``
prompt template. The prompt now says "decide the count yourself based
on what you see — the smallest set that still covers every recurring
phase" and adds an "each label must recur across the demos" rule so
the VLM filters out one-off motions.
Update the launcher script + docs to remove the old knobs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Three stale things in the launcher script:
- ``--module_1/2/3.*`` no longer exist; review commit fd18beb renamed
the CLI namespaces to ``--plan/interjections/vqa``. Forwarded all
eight existing args to their new names.
- ``--push_to_hub`` is now a bool; the destination repo lives at
``--dest_repo_id``. Split the single positional into both args.
- ``openai`` was missing from the pip install list, which the prior
review review (claude bot, 2026-05-08) flagged — the default vlm
backend is ``openai`` so the job would have ImportError'd. Added.
Also expose the new phase 0 (canonical vocabulary discovery) knobs
explicitly: ``--vocabulary.sample_episodes``, ``--n_subtask_target``,
``--n_memory_target``. Defaults are sane (3 / 10 / 6) but worth
flagging in the example so the operator knows what they're running.
Update the docstring + section comments to match the current phase
layout (vocabulary → plan → interjections → vqa → writer).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Adds an optional `dest_repo_id` to AnnotationPipelineConfig. When set,
`push_to_hub` uploads the annotated dataset there instead of overwriting
the source `repo_id`, restoring separate source/destination repos.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- name the three modules everywhere (plan / interjections / vqa) instead
of module_1/2/3 — config classes, config fields, executor params,
staging keys and phase names now carry the module name
- rename examples/annotation -> examples/annotations; add the Apache
header to run_hf_job.py
- drop the unused GeneralVqaModule._generate_one
- remove "PR 1" references from comments/docstrings
- frames.py: rely on the always-defined LeRobotDatasetMetadata.camera_keys
- executor.py: read/write meta/info.json via load_info / write_info
- reader.py: load meta/tasks.parquet via io_utils.load_tasks
- make --push_to_hub a bool; push the annotated dataset back to --repo_id
- move the on-disk test dataset builder into tests/fixtures
(build_annotation_dataset); run_e2e_smoke reuses it
- clarify in the docs that the vqa module grounds each pair on a single
frame (K = per-tick anchor count)
- hoist stdlib dynamic imports to module scope
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
