fix typo

Co-authored-by: Martino Russi <77496684+nepyope@users.noreply.github.com>

docs/source/annotation_pipeline.mdx

@@ -81,6 +81,12 @@ merged. Both prompts also carry a causal **event-boundary** definition (a
 new event starts when an object becomes held / is released / reaches a new
 location / a lid changes state / contents move) to sharpen where cuts land.
 
+Optionally, a third **seeded-relabel** pass (`--plan.subtask_seeded_relabel`)
+revisits each span with its previous/current/next segment contact sheets and
+minimally corrects the label, using the first label as a prior — it keeps the
+boundaries fixed and only sharpens wording, at the cost of one extra call per
+subtask.
+
 The resulting spans are then stitched into a gap-free, full-episode
 cover, so **every frame has exactly one active subtask**. See
 [`run_hf_job.py`](https://github.com/huggingface/lerobot/blob/main/examples/annotations/run_hf_job.py)
@@ -157,30 +163,33 @@ Every module is on by default and can be toggled independently (set to
 
 ### The VLM (`--vlm.*`)
 
-| Flag                       | Default            | What it does                                                                        |
-| -------------------------- | ------------------ | ----------------------------------------------------------------------------------- |
-| `--vlm.model_id`           | `Qwen/Qwen3.6-27B` | The model to serve and prompt.                                                      |
-| `--vlm.camera_key`         | first `images.*`   | Which camera every prompt is grounded on.                                           |
-| `--vlm.serve_command`      | auto               | The exact `vllm serve …` command (set TP size, GPU memory, `--max-model-len` here). |
-| `--vlm.parallel_servers`   | `1`                | Independent servers for round-robin routing (one per GPU).                          |
-| `--vlm.num_gpus`           | `0`                | GPUs per server (`0` = one each).                                                   |
-| `--vlm.client_concurrency` | `16`               | In-flight requests across all servers.                                              |
-| `--vlm.max_new_tokens`     | `512`              | Generation cap per call.                                                            |
-| `--vlm.temperature`        | `0.2`              | Sampling temperature.                                                               |
+| Flag                       | Default            | What it does                                                                         |
+| -------------------------- | ------------------ | ------------------------------------------------------------------------------------ |
+| `--vlm.model_id`           | `Qwen/Qwen3.6-27B` | The model to serve and prompt.                                                       |
+| `--vlm.camera_key`         | first `images.*`   | Which camera every prompt is grounded on.                                            |
+| `--vlm.serve_command`      | auto               | The exact `vllm serve …` command (set TP size, GPU memory, `--max-model-len` here).  |
+| `--vlm.parallel_servers`   | `1`                | Independent servers for round-robin routing (one per GPU).                           |
+| `--vlm.num_gpus`           | `0`                | GPUs per server (`0` = one each).                                                    |
+| `--vlm.client_concurrency` | `16`               | In-flight requests across all servers.                                               |
+| `--vlm.max_new_tokens`     | `512`              | Generation cap per call.                                                             |
+| `--vlm.temperature`        | `0.2`              | Sampling temperature.                                                                |
+| `--vlm.reasoning_effort`   | `null`             | Thinking-budget hint (`low`/`medium`/`high`) forwarded to OpenAI-compatible servers. |
 
 ### Subtasks / plan / memory (`--plan.*`)
 
-| Flag                            | Default    | What it does                                                                                                              |
-| ------------------------------- | ---------- | ------------------------------------------------------------------------------------------------------------------------- |
-| `--plan.frames_per_second`      | `2.0`      | Frame sampling rate for the contact sheets (`2.0` = one frame every 0.5s).                                                |
-| `--plan.max_frames_per_prompt`  | `60`       | Frame budget per VLM call. Episodes whose sampling exceeds this are auto-windowed at the same density, then stitched.     |
-| `--plan.contact_sheet_columns`  | `5`        | Columns per contact-sheet grid (`contact_sheet_frames_per_sheet` tiles, time row-major).                                  |
-| `--plan.plan_max_steps`         | `8`        | Upper bound on subtasks per episode.                                                                                      |
-| `--plan.subtask_describe_first` | `true`     | Run the describe→segment grounding pass (best subtask quality; +1 call/episode).                                          |
-| `--plan.emit_plan`              | `true`     | Emit the numbered `plan` rows (`false` = subtasks + memory only).                                                         |
-| `--plan.emit_memory`            | `true`     | Emit the `memory` rows (`false` = subtasks + plan only); symmetric to `emit_plan`.                                        |
-| `--plan.n_task_rephrasings`     | `10`       | How many `task_aug` rephrasings to emit (`0` disables).                                                                   |
-| `--plan.derive_task_from_video` | `if_short` | Use the dataset task as-is (`off`), only when it's missing/short (`if_short`), or always re-derive from video (`always`). |
+| Flag                            | Default    | What it does                                                                                                                 |
+| ------------------------------- | ---------- | ---------------------------------------------------------------------------------------------------------------------------- |
+| `--plan.frames_per_second`      | `2.0`      | Frame sampling rate for the contact sheets (`2.0` = one frame every 0.5s).                                                   |
+| `--plan.max_frames_per_prompt`  | `60`       | Frame budget per VLM call. Episodes whose sampling exceeds this are auto-windowed at the same density, then stitched.        |
+| `--plan.contact_sheet_columns`  | `5`        | Columns per contact-sheet grid (`contact_sheet_frames_per_sheet` tiles, time row-major).                                     |
+| `--plan.plan_max_steps`         | `8`        | Upper bound on subtasks per episode.                                                                                         |
+| `--plan.subtask_describe_first` | `true`     | Run the describe→segment grounding pass (best subtask quality; +1 call/episode).                                             |
+| `--plan.subtask_seeded_relabel` | `false`    | Second pass: re-label each subtask from its prev/current/next contact sheets, seeded with the first label (+1 call/subtask). |
+| `--plan.subtask_relabel_frames` | `5`        | Frames sampled uniformly per segment sheet in the relabel pass (only used when `subtask_seeded_relabel=true`).               |
+| `--plan.emit_plan`              | `true`     | Emit the numbered `plan` rows (`false` = subtasks + memory only).                                                            |
+| `--plan.emit_memory`            | `true`     | Emit the `memory` rows (`false` = subtasks + plan only); symmetric to `emit_plan`.                                           |
+| `--plan.n_task_rephrasings`     | `10`       | How many `task_aug` rephrasings to emit (`0` disables).                                                                      |
+| `--plan.derive_task_from_video` | `if_short` | Use the dataset task as-is (`off`), only when it's missing/short (`if_short`), or always re-derive from video (`always`).    |
 
 ### Interjections + VQA
 

examples/annotations/run_hf_job.py

@@ -46,8 +46,11 @@ CMD = (
     "apt-get update -qq && apt-get install -y -qq git ffmpeg && "
     "pip install --no-deps "
     "'lerobot @ git+https://github.com/huggingface/lerobot.git@main' && "
+    # Pins mirror pyproject.toml — unpinned installs pull av 18 / datasets 5 /
+    # draccus 0.11, which break lerobot at import time.
     "pip install --upgrade-strategy only-if-needed "
-    "datasets pyarrow av jsonlines draccus gymnasium torchcodec mergedeep pyyaml-include toml typing-inspect "
+    "'datasets>=4.7.0,<5.0.0' 'pyarrow>=21.0.0,<30.0.0' 'av>=15.0.0,<16.0.0' 'draccus==0.10.0' "
+    "'pandas>=2.0.0,<3.0.0' jsonlines gymnasium torchcodec mergedeep pyyaml-include toml typing-inspect "
     "openai && "
     "export VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0 && "
     "export VLLM_VIDEO_BACKEND=pyav && "

src/lerobot/annotations/steerable_pipeline/config.py

@@ -65,6 +65,14 @@ class PlanConfig:
     # invented from the task text (+1 VLM call/episode).
     subtask_describe_first: bool = True
 
+    # Seeded relabeling: after segmentation, re-label each span with a focused
+    # pass that sees the previous / current / next segment contact sheets and
+    # minimally corrects the seed label (macrodata's best end-to-end labeling
+    # step). Costs +1 VLM call per subtask; off by default.
+    subtask_seeded_relabel: bool = False
+    # Frames sampled uniformly per segment sheet in the relabel pass.
+    subtask_relabel_frames: int = 5
+
     # Emit ``style="plan"`` rows at each boundary; False = subtasks + memory only.
     emit_plan: bool = True
 
@@ -160,6 +168,11 @@ class VlmConfig:
     # Forwarded as extra_body.chat_template_kwargs (e.g. {"enable_thinking": false}).
     chat_template_kwargs: dict[str, Any] | None = None
 
+    # OpenAI-style thinking budget hint ("low"/"medium"/"high"); forwarded to
+    # the server when set. Used to cap a thinking model's reasoning so it
+    # leaves tokens for the actual JSON answer on OpenAI-compatible endpoints.
+    reasoning_effort: str | None = None
+
 
 @dataclass
 class ExecutorConfig:

src/lerobot/annotations/steerable_pipeline/frames.py

@@ -413,7 +413,16 @@ def _draw_timestamp_badge(image: PIL.Image.Image, timestamp: float) -> PIL.Image
 
     result = image.copy()
     draw = ImageDraw.Draw(result)
-    font = ImageFont.load_default()
+    # Scale the timestamp to the tile so it stays legible after the model
+    # downsamples the full sheet into 768px tiles — a tiny bitmap font blurs
+    # at contact-sheet resolution and the VLM can no longer read the exact
+    # source time, which is what the boundary score depends on. ``size=`` is
+    # supported by Pillow's bitmap default since 10.1; fall back otherwise.
+    badge_px = max(14, round(image.height * 0.12))
+    try:
+        font = ImageFont.load_default(size=badge_px)
+    except TypeError:
+        font = ImageFont.load_default()
     label = f"{timestamp:06.2f}s"
     left, top, right, bottom = draw.textbbox((0, 0), label, font=font)
     text_w, text_h = right - left, bottom - top

src/lerobot/annotations/steerable_pipeline/modules/plan_subtasks_memory.py

@@ -116,6 +116,8 @@ class PlanSubtasksMemoryModule:
             rows.extend(self._task_aug_rows([effective_task, *variants], t0))
 
         subtask_spans = self._generate_subtasks(record, task=effective_task)
+        if self.config.subtask_seeded_relabel and subtask_spans:
+            subtask_spans = self._seeded_relabel(record, subtask_spans, effective_task)
 
         # subtask rows
         for span in subtask_spans:
@@ -509,6 +511,51 @@ class PlanSubtasksMemoryModule:
 
         return cleaned
 
+    def _seeded_relabel(
+        self, record: EpisodeRecord, spans: list[dict[str, Any]], task: str
+    ) -> list[dict[str, Any]]:
+        """Re-label each span using prev/current/next segment contact sheets.
+
+        Boundaries are kept fixed; only ``text`` is refined. The original
+        ("seed") label is passed as a strong prior so the model verifies and
+        minimally corrects it rather than re-describing from scratch — the
+        macrodata seeded-relabeling step. One VLM call per span.
+        """
+        n = len(spans)
+        out: list[dict[str, Any]] = []
+        for i, span in enumerate(spans):
+            content: list[dict[str, Any]] = []
+            if i > 0:
+                content += self._segment_sheet(record, spans[i - 1])
+            content += self._segment_sheet(record, span)
+            if i < n - 1:
+                content += self._segment_sheet(record, spans[i + 1])
+            prompt = load_prompt("plan_subtask_relabel").format(
+                episode_task=task,
+                seed_label=span["text"],
+                segment_index=i + 1,
+                segment_count=n,
+                start=float(span["start"]),
+                end=float(span["end"]),
+            )
+            content.append({"type": "text", "text": prompt})
+            label = self._vlm_field([{"role": "user", "content": content}], "label")
+            text = label.strip() if isinstance(label, str) and label.strip() else span["text"]
+            out.append({**span, "text": text})
+        return out
+
+    def _segment_sheet(self, record: EpisodeRecord, span: dict[str, Any]) -> list[dict[str, Any]]:
+        """Contact-sheet block(s) for one span: up to N frames sampled uniformly."""
+        s, e = float(span["start"]), float(span["end"])
+        n = max(1, int(self.config.subtask_relabel_frames))
+        if e <= s or n == 1:
+            timestamps = [s]
+        else:
+            step = (e - s) / (n - 1)
+            timestamps = [s + i * step for i in range(n)]
+        frames = self.frame_provider.frames_at(record, timestamps)
+        return self._contact_sheet_blocks(frames, timestamps[: len(frames)])
+
     def _generate_subtasks_windowed(
         self, record: EpisodeRecord, task: str, window_s: float
     ) -> list[dict[str, Any]]:

src/lerobot/annotations/steerable_pipeline/prompts/__init__.py

@@ -22,12 +22,23 @@ plain editors and roundtrip cleanly through ``ruff format``.
 
 from __future__ import annotations
 
+import os
 from pathlib import Path
 
 _DIR = Path(__file__).parent
 
 
 def load(name: str) -> str:
-    """Read prompt template ``name.txt`` from the ``prompts/`` directory."""
+    """Read prompt template ``name.txt`` from the ``prompts/`` directory.
+
+    A ``LEROBOT_PROMPT_OVERRIDE_<name>`` environment variable, when set to a
+    non-empty value, takes precedence over the packaged file. This lets prompt
+    search (e.g. GEPA) inject candidate templates into a remote job without
+    rebuilding the package; the override must keep the same ``{placeholder}``
+    fields the call site formats in.
+    """
+    override = os.environ.get(f"LEROBOT_PROMPT_OVERRIDE_{name}")
+    if override and override.strip():
+        return override
     path = _DIR / f"{name}.txt"
     return path.read_text(encoding="utf-8")

src/lerobot/annotations/steerable_pipeline/prompts/plan_subtask_relabel.txt

@@ -0,0 +1,35 @@
+Annotate one fixed segment from a longer robot demonstration.
+
+Return only JSON:
+  {{"label": "<short descriptive subtask label>"}}
+
+You are shown up to three timestamped contact sheets, in order:
+- The FIRST sheet is the PREVIOUS segment (context only); it may be absent.
+- The SECOND sheet is the CURRENT target segment.
+- The THIRD sheet is the NEXT segment (context only); it may be absent.
+Each tile has its timestamp (seconds, absolute video time) burned into its
+top-left corner.
+
+Episode instruction: "{episode_task}"
+Target segment: {segment_index} of {segment_count}
+Target time: {start:.2f}s to {end:.2f}s
+Original predicted label for this exact segment: "{seed_label}"
+
+Rules:
+- Label ONLY the current target segment (the second sheet). Use the
+  previous/next sheets only to disambiguate what changed.
+- Treat the original predicted label as a STRONG PRIOR, not ground truth:
+  verify it against the current segment and correct it minimally.
+- If it already names the right action and main object, keep it; only fix
+  grammar or add a clearly visible essential detail.
+- If it is vague but directionally correct, make it more specific.
+- If it describes the previous/next segment, the wrong action, wrong
+  object, wrong destination, or a wrong state change, replace it.
+- Do not describe the previous or next segment, and do not split, merge,
+  or move the fixed segment.
+- Do not introduce an action that is not clearly visible in the current
+  target segment.
+- Use one concise imperative phrase. Name the manipulated object and the
+  action / state change. Include source, destination, side, direction,
+  final placement, or opened/closed state when visible and central.
+- Do not mention timestamps, frame numbers, uncertainty, or intent.

src/lerobot/annotations/steerable_pipeline/prompts/plan_subtasks.txt

@@ -1,112 +1,68 @@
-You are labeling a teleoperated robot demonstration.
+You are annotating a teleoperated robot demonstration shown as
+timestamped contact sheets (each tile has its time in seconds burned
+into the top-left corner). The operator's goal was: "{episode_task}"
 
-The user originally asked: "{episode_task}"
+{observation_block}Reconstruct the sequence of COMPLETED manipulation events the robot
+performs, in chronological order. Output one segment per event with a
+[start, end] time in seconds and a short action label.
 
-You are shown the entire demonstration as a single video. Watch the
-whole clip, then segment it into a list of consecutive atomic subtasks
-the robot performs.
+GROUNDING — read first, it overrides everything below:
+- Label ONLY events you can SEE in the frames. The instruction is the
+  goal; the VIDEO is the ground truth for what actually happened.
+- Do NOT invent, anticipate, or pad steps that are not shown.
 
-{observation_block}GROUNDING — read this first, it overrides everything below:
-- Label ONLY what the robot actually does in the video. Every subtask
-  you emit must correspond to motion you can SEE in specific frames.
-- Do NOT invent, anticipate, or pad. If the robot only does one thing
-  (e.g. it just navigates to a location and the clip ends), emit
-  EXACTLY ONE subtask. Many demonstrations are a single atomic skill.
-- ``max_steps`` below is a hard CEILING, not a target. Emitting fewer
-  subtasks than the ceiling is not just allowed, it is expected for
-  short / atomic demonstrations. One correct subtask is far better
-  than several invented ones.
-- If the video does not clearly show the action implied by the task,
-  describe what you actually see — do NOT fabricate the task's steps
-  from the instruction text. The instruction tells you the goal; the
-  VIDEO is the ground truth for what happened.
+Granularity — segment by completed events, not by motion:
+- Start a NEW segment whenever the world state changes: an object is
+  grasped, lifted, transported, placed, or released; a held object
+  changes; a drawer/door/lid/container opens or closes; contents move
+  between containers (poured); a tool starts or stops acting on a
+  surface. Watch the gripper open/close transitions — they usually mark
+  boundaries.
+- Do NOT split approach, reach, grasp adjustment, small repositioning,
+  hesitation, or retreat into their own segments. Fold each into the
+  event it belongs to (the approach is part of the pick; the retreat is
+  part of the place).
+- Do NOT merge separate completed events. Each distinct pick, place,
+  open, close, pour, push, wipe, or insert is its own segment, even when
+  they repeat on different objects or locations.
+- Most segments last 2-10 seconds. Shorter segments are okay ONLY for
+  fast pick / place / open / close / release events. Never emit a
+  segment shorter than {min_subtask_seconds} seconds; merge a too-short
+  candidate into its neighbour instead.
+- Skip idle time, pure camera motion, and tiny hand jitter.
 
-Authoring rules — Hi Robot atom granularity, pi0.7-style short prompts:
+Labels — short imperative phrases:
+- One concise command naming the action and the manipulated object, e.g.
+  "pick up the red cup", "put the cup on the shelf", "open the top
+  drawer", "pour water into the glass", "insert the plug into the
+  socket".
+- Include source, destination, side, direction, or the final
+  open/closed state when it is visible and central to the event.
+- Prefer these verbs (extend only when none fits): pick up, put, place,
+  push, pull, turn, press, open, close, pour, insert, wipe, stack.
+  Disambiguate by what you SEE:
+    * STACK vs PUT: object placed ON TOP OF another object -> "stack".
+    * INSERT vs PUT: object pushed INTO a fitted slot/hole/socket -> "insert".
+    * PICK UP vs PUT (direction): gripper CLOSES and object moves WITH
+      the hand -> "pick up"; gripper OPENS and object stays -> "put".
+    * POUR vs PUT: source is tilted and contents flow -> "pour".
+- Use the exact object nouns implied by the task; stay consistent across
+  the episode (don't switch "cube" to "block").
+- Write imperative commands, never third person ("the robot ..."), and
+  drop articles/adverbs.
 
-- Each subtask = one COMPOSITE atomic skill the low-level policy can
-  execute end-to-end. A "skill" bundles its own approach motion with
-  its terminal action — do NOT split the approach off as its own
-  subtask. The whole-arm policy already learns to reach as part of
-  every manipulation primitive.
-- Write each subtask as an IMPERATIVE COMMAND, starting with one of
-  these verbs (extend only when none fits):
-    pick up <obj>           — approach + grasp + lift in one subtask
-    put <obj> on/in <loc>   — transport + release in one subtask
-    place <obj> on/in <loc> — synonym of "put"; pick one and stay consistent
-    push <obj>              — contact + linear shove
-    pull <obj>              — contact + linear retract
-    turn <knob/dial/handle> — rotary actuation
-    press <button>          — single-press contact
-    open <drawer/door/lid>  — full open motion
-    close <drawer/door/lid> — full close motion
-    pour <src> into <dst>   — tilt + flow
-    insert <obj> into <slot>— alignment + push-fit
-    go to <loc>             — ONLY when no grasp / actuation follows
-                             (e.g. a pure relocation between phases).
-                             If the next subtask grasps something at
-                             that location, drop "go to ..." and just
-                             write "pick up ..." instead.
-- Forbidden ultra-fine splits — the VLM is NOT allowed to emit these
-  as standalone subtasks; fold them into the parent composite:
-    "move to X"   → fold into "pick up X" (or whatever follows)
-    "reach for X" → fold into "pick up X"
-    "grasp X"     → fold into "pick up X"
-    "lift X"      → fold into "pick up X" (or "put X on Y" if it's
-                    the transport phase of a place)
-    "release X"   → fold into "put X on Y" (or "place X in Y")
-- Keep it SHORT — a verb phrase, not a sentence. Drop articles
-  ("the", "a") and adverbs ("carefully", "slowly"). Add a "how"
-  detail (which hand, which grasp point) ONLY when it is needed to
-  disambiguate. Every subtask must begin with one of the verbs
-  above (no leading nouns, no "then", no "first").
-- NEVER use third person. Never write "the robot", "the arm", "the
-  gripper moves", "it picks up" — the robot is implied. Command it,
-  do not describe it.
-- Use the exact object nouns from the task above. If the task says
-  "cube", every subtask says "cube" — never switch to "block". If it
-  says "box", never switch to "bin"/"container". Keep vocabulary
-  consistent across the whole episode.
-- Good: "pick up blue cube", "put blue cube in box", "open drawer",
-  "turn red knob", "press start button", "go to sink".
-- Bad: "move to blue cube" (approach as its own subtask — forbidden,
-  must be folded into "pick up blue cube"); "the robot arm moves
-  towards the blue cube" (third person, too long); "carefully pick
-  up the cube" (adverb, article); "release the yellow block"
-  ("block" when the task said "cube", and "release" must be folded
-  into a "put"/"place" subtask).
-- Subtasks are non-overlapping and cover the full episode in order.
-  Choose the cut points yourself based on what you see in the video
-  (gripper open/close events, contact, regrasps, transitions).
-- Each subtask spans at least {min_subtask_seconds} seconds. If a
-  candidate span would be shorter, merge it into its neighbour
-  rather than emitting it.
-- Do not exceed {max_steps} subtasks total. Fewer, larger composites
-  are preferred over many micro-steps.
-- Every subtask's [start_time, end_time] must lie within
-  [0.0, {episode_duration}] seconds.
-
-SPECIAL CASES — verb disambiguation (each rule is narrowly visual and
-fires ONLY on the spatial situation it names; it must not change how you
-label any other situation):
-- STACK vs PUT: if an object is placed ON TOP OF another specific object
-  (not on a flat table / shelf / counter), use "stack ... on ...", not
-  "put". "stack blue book on green book", NOT "put blue book on table".
-- INSERT vs PUT: if an object goes INTO a fitted slot / hole / socket /
-  receptacle (push-fit), use "insert ... into ...", not "put".
-- RETRIEVE/PICK-UP vs PUT (direction): watch the gripper. If it CLOSES
-  on the object and the object moves WITH the hand, it is "pick up" /
-  "retrieve" (object leaves its location). If the gripper OPENS and the
-  object stays where the hand left it, it is "put" / "place" (object
-  arrives at a location). Decide by which way the object moves, not by
-  where the hand ends up.
-- POUR vs PUT: only use "pour" when the source is tilted and contents
-  flow out; moving a full container without tilting is "put"/"place".
+Timing:
+- Use the burned-in timestamps to set start and end. Boundaries should
+  land on or near a printed time, and every [start, end] must lie within
+  [0.0, {episode_duration}] seconds, be non-overlapping, and cover the
+  episode in order.
+- Emit at most {max_steps} segments.
 
 Output strictly valid JSON of shape:
 
   {{
     "subtasks": [
-      {{"text": "<short imperative verb phrase>", "start": <float>, "end": <float>}},
+      {{"text": "<short imperative action label>", "start": <float>, "end": <float>}},
       ...
     ]
   }}

src/lerobot/annotations/steerable_pipeline/vlm_client.py

@@ -285,6 +285,8 @@ def _make_openai_client(config: VlmConfig) -> VlmClient:
             "max_tokens": max_tok,
             "temperature": temp,
         }
+        if config.reasoning_effort:
+            kwargs["reasoning_effort"] = config.reasoning_effort
         extra_body: dict[str, Any] = {}
         if send_mm_kwargs and mm_kwargs:
             extra_body["mm_processor_kwargs"] = {**mm_kwargs, "do_sample_frames": True}
@@ -296,7 +298,13 @@ def _make_openai_client(config: VlmConfig) -> VlmClient:
             chosen = clients[rr_counter["i"] % len(clients)]
             rr_counter["i"] += 1
         response = chosen.chat.completions.create(**kwargs)
-        return response.choices[0].message.content or ""
+        # Some OpenAI-compatible servers can return a choice with no message
+        # (safety filter, or a "thinking" model that spends the whole budget
+        # before emitting content). Treat that as an empty reply so the
+        # JSON-retry path handles it instead of crashing the run.
+        choice = response.choices[0] if response.choices else None
+        message = choice.message if choice is not None else None
+        return (message.content if message is not None else None) or ""
 
     def _gen(batch: Sequence[Sequence[dict[str, Any]]], max_tok: int, temp: float) -> list[str]:
         if len(batch) <= 1 or config.client_concurrency <= 1:

src/lerobot/policies/evo1/internvl3_embedder.py

@@ -27,9 +27,11 @@ from lerobot.utils.import_utils import _transformers_available, require_package
 
 if TYPE_CHECKING or _transformers_available:
     from transformers import AutoModel, AutoTokenizer
+    from transformers.utils import is_flash_attn_2_available
 else:
     AutoModel = None
     AutoTokenizer = None
+    is_flash_attn_2_available = None
 
 IMAGENET_MEAN = (0.485, 0.456, 0.406)
 IMAGENET_STD = (0.229, 0.224, 0.225)
@@ -135,9 +137,13 @@ class InternVL3Embedder(nn.Module):
                 raise ValueError(f"Unsupported EVO1 vlm_dtype '{model_dtype}'") from exc
         self.model_dtype = model_dtype
 
-        attn_implementation = "flash_attention_2" if (use_flash_attn and _flash_attn_available()) else "eager"
+        attn_implementation = (
+            "flash_attention_2" if (use_flash_attn and is_flash_attn_2_available()) else "eager"
+        )
         if use_flash_attn and attn_implementation == "eager":
-            logger.warning("flash_attn is not installed. Falling back to eager attention.")
+            logger.warning(
+                "Flash Attention 2 is unavailable on this runtime. Falling back to eager attention."
+            )
 
         self.model = AutoModel.from_pretrained(
             model_name,
@@ -359,11 +365,3 @@ class InternVL3Embedder(nn.Module):
     @property
     def device(self) -> torch.device:
         return next(self.model.parameters()).device
-
-
-def _flash_attn_available() -> bool:
-    try:
-        import flash_attn  # noqa: F401
-    except ModuleNotFoundError:
-        return False
-    return True

src/lerobot/policies/pretrained.py

@@ -23,8 +23,6 @@ from pathlib import Path
 from tempfile import TemporaryDirectory
 from typing import TYPE_CHECKING, TypedDict, TypeVar, Unpack
 
-import packaging
-import safetensors
 from huggingface_hub import HfApi, ModelCard, ModelCardData, hf_hub_download, save_torch_state_dict
 from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
 from huggingface_hub.errors import HfHubHTTPError
@@ -34,6 +32,7 @@ from torch import Tensor, nn
 from lerobot.__version__ import __version__
 from lerobot.configs import PreTrainedConfig
 from lerobot.configs.train import TrainPipelineConfig
+from lerobot.utils.device_utils import resolve_safetensors_device
 from lerobot.utils.hub import HubMixin
 
 from .utils import log_model_loading_keys
@@ -221,26 +220,10 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
 
     @classmethod
     def _load_as_safetensor(cls, model: T, model_file: str, map_location: str, strict: bool) -> T:
-        # Create base kwargs
-        kwargs = {"strict": strict}
-
-        # Add device parameter for newer versions that support it
-        if packaging.version.parse(safetensors.__version__) >= packaging.version.parse("0.4.3"):
-            kwargs["device"] = map_location
-
-        # Load the model with appropriate kwargs
-        missing_keys, unexpected_keys = load_model_as_safetensor(model, model_file, **kwargs)
+        missing_keys, unexpected_keys = load_model_as_safetensor(
+            model, model_file, strict=strict, device=resolve_safetensors_device(map_location)
+        )
         log_model_loading_keys(missing_keys, unexpected_keys)
-
-        # For older versions, manually move to device if needed
-        if "device" not in kwargs and map_location != "cpu":
-            logging.warning(
-                "Loading model weights on other devices than 'cpu' is not supported natively in your version of safetensors."
-                " This means that the model is loaded on 'cpu' first and then copied to the device."
-                " This leads to a slower loading time."
-                " Please update safetensors to version 0.4.3 or above for improved performance."
-            )
-            model.to(map_location)
         return model
 
     @abc.abstractmethod

src/lerobot/policies/xvla/modeling_xvla.py

@@ -154,7 +154,7 @@ class XVLAModel(nn.Module):
         # Freeze or unfreeze policy transformer
         if not self.config.train_policy_transformer:
             for name, param in self.transformer.named_parameters():
-                if "soft_prompts" not in name:
+                if "soft_prompt" not in name:
                     param.requires_grad = False
 
         # Freeze or unfreeze soft prompts

src/lerobot/rewards/pretrained.py

@@ -21,8 +21,6 @@ from pathlib import Path
 from tempfile import TemporaryDirectory
 from typing import TYPE_CHECKING, Any, TypeVar
 
-import packaging
-import safetensors
 from huggingface_hub import HfApi, ModelCard, ModelCardData, hf_hub_download
 from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
 from huggingface_hub.errors import HfHubHTTPError
@@ -30,6 +28,7 @@ from safetensors.torch import load_model as load_model_as_safetensor, save_model
 from torch import Tensor, nn
 
 from lerobot.configs.rewards import RewardModelConfig
+from lerobot.utils.device_utils import resolve_safetensors_device
 from lerobot.utils.hub import HubMixin
 
 if TYPE_CHECKING:
@@ -129,29 +128,13 @@ class PreTrainedRewardModel(nn.Module, HubMixin, abc.ABC):
 
     @classmethod
     def _load_as_safetensor(cls, model: T, model_file: str, map_location: str, strict: bool) -> T:
-        # Create base kwargs
-        kwargs = {"strict": strict}
-
-        # Add device parameter for newer versions that support it
-        if packaging.version.parse(safetensors.__version__) >= packaging.version.parse("0.4.3"):
-            kwargs["device"] = map_location
-
-        # Load the model with appropriate kwargs
-        missing_keys, unexpected_keys = load_model_as_safetensor(model, model_file, **kwargs)
+        missing_keys, unexpected_keys = load_model_as_safetensor(
+            model, model_file, strict=strict, device=resolve_safetensors_device(map_location)
+        )
         if missing_keys:
             logging.warning(f"Missing key(s) when loading model: {missing_keys}")
         if unexpected_keys:
             logging.warning(f"Unexpected key(s) when loading model: {unexpected_keys}")
-
-        # For older versions, manually move to device if needed
-        if "device" not in kwargs and map_location != "cpu":
-            logging.warning(
-                "Loading model weights on other devices than 'cpu' is not supported natively in your version of safetensors."
-                " This means that the model is loaded on 'cpu' first and then copied to the device."
-                " This leads to a slower loading time."
-                " Please update safetensors to version 0.4.3 or above for improved performance."
-            )
-            model.to(map_location)
         return model
 
     def get_optim_params(self):

src/lerobot/scripts/lerobot_annotate.py

@@ -28,7 +28,12 @@ For distributed runs, see ``examples/annotations/run_hf_job.py``.
 """
 
 import logging
+from contextlib import suppress
 from pathlib import Path
+from typing import TYPE_CHECKING
+
+from huggingface_hub import HfApi, snapshot_download
+from huggingface_hub.errors import RevisionNotFoundError
 
 from lerobot.annotations.steerable_pipeline.config import AnnotationPipelineConfig
 from lerobot.annotations.steerable_pipeline.executor import Executor
@@ -42,6 +47,12 @@ from lerobot.annotations.steerable_pipeline.validator import StagingValidator
 from lerobot.annotations.steerable_pipeline.vlm_client import make_vlm_client
 from lerobot.annotations.steerable_pipeline.writer import LanguageColumnsWriter
 from lerobot.configs import parser
+from lerobot.utils.import_utils import _datasets_available, require_package
+
+if TYPE_CHECKING or _datasets_available:
+    from lerobot.datasets.dataset_metadata import CODEBASE_VERSION
+    from lerobot.datasets.io_utils import load_info
+    from lerobot.datasets.utils import create_lerobot_dataset_card
 
 logger = logging.getLogger(__name__)
 
@@ -50,8 +61,6 @@ def _resolve_root(cfg: AnnotationPipelineConfig) -> Path:
     if cfg.root is not None:
         return Path(cfg.root)
     if cfg.repo_id is not None:
-        from huggingface_hub import snapshot_download
-
         return Path(snapshot_download(repo_id=cfg.repo_id, repo_type="dataset"))
     raise ValueError("Either --root or --repo_id must be provided.")
 
@@ -125,7 +134,7 @@ def _push_to_hub(root: Path, cfg: AnnotationPipelineConfig) -> None:
 
     Pushes to ``cfg.new_repo_id`` when set, otherwise back to ``cfg.repo_id``.
     """
-    from huggingface_hub import HfApi  # noqa: PLC0415
+    require_package("datasets", "dataset")
 
     repo_id = cfg.new_repo_id or cfg.repo_id
     commit_message = cfg.push_commit_message or "Add steerable annotations (lerobot-annotate)"
@@ -143,33 +152,26 @@ def _push_to_hub(root: Path, cfg: AnnotationPipelineConfig) -> None:
         repo_id=repo_id,
         repo_type="dataset",
         commit_message=commit_message,
-        ignore_patterns=[".annotate_staging/**", "**/.DS_Store"],
+        # README.md is excluded because when pushing to ``new_repo_id`` the
+        # source card's links (e.g. the visualize badge) would keep pointing
+        # at the source dataset; a fresh card is generated below instead.
+        ignore_patterns=[".annotate_staging/**", "**/.DS_Store", "README.md"],
     )
     print(f"[lerobot-annotate] uploaded to https://huggingface.co/datasets/{repo_id}", flush=True)
 
+    dataset_info = load_info(root)
+    card = create_lerobot_dataset_card(dataset_info=dataset_info, license="apache-2.0", repo_id=repo_id)
+    card.push_to_hub(repo_id=repo_id, repo_type="dataset")
+
     # Tag the upload with the codebase version. ``LeRobotDatasetMetadata``
     # resolves the dataset revision via ``get_safe_version`` which scans
     # for tags like ``v3.0``; without a tag it raises
     # ``RevisionNotFoundError``. Read the version straight from the
     # dataset's own ``meta/info.json`` so we tag whatever the writer
     # actually wrote (no accidental drift if the codebase floor moves).
-    from lerobot.datasets.dataset_metadata import CODEBASE_VERSION  # noqa: PLC0415
-
-    info_path = root / "meta" / "info.json"
-    version_tag = CODEBASE_VERSION
-    if info_path.exists():
-        try:
-            from lerobot.utils.io_utils import load_json  # noqa: PLC0415
-
-            info = load_json(info_path)
-            ds_version = info.get("codebase_version")
-            if isinstance(ds_version, str) and ds_version.startswith("v"):
-                version_tag = ds_version
-        except Exception as exc:  # noqa: BLE001
-            print(
-                f"[lerobot-annotate] could not read codebase_version from info.json ({exc}); falling back to {version_tag}",
-                flush=True,
-            )
+    version_tag = (
+        dataset_info.codebase_version if dataset_info.codebase_version.startswith("v") else CODEBASE_VERSION
+    )
     revision = getattr(commit_info, "oid", None)
     tag_kwargs = {
         "repo_id": repo_id,
@@ -180,10 +182,6 @@ def _push_to_hub(root: Path, cfg: AnnotationPipelineConfig) -> None:
         tag_kwargs["revision"] = revision
 
     try:
-        from contextlib import suppress  # noqa: PLC0415
-
-        from huggingface_hub.errors import RevisionNotFoundError  # noqa: PLC0415
-
         with suppress(RevisionNotFoundError):
             api.delete_tag(repo_id, tag=version_tag, repo_type="dataset")
         api.create_tag(**tag_kwargs)

src/lerobot/scripts/lerobot_train.py

@@ -171,6 +171,9 @@ def update_policy(
     train_metrics.update_s = time.perf_counter() - start_time
     if torch.cuda.is_available():
         train_metrics.gpu_mem_gb = torch.cuda.max_memory_allocated() / (1024**3)
+    # Aggregate the policy's scalar outputs for logging and rank-reduction across the log window.
+    if output_dict:
+        train_metrics.update_metrics(output_dict)
     return train_metrics, output_dict
 
 
@@ -572,7 +575,7 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         batch = preprocessor(batch)
         train_tracker.dataloading_s = time.perf_counter() - start_time
 
-        train_tracker, output_dict = update_policy(
+        train_tracker, _ = update_policy(
             train_tracker,
             policy,
             batch,
@@ -605,9 +608,10 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
                     train_tracker.samples_per_s = effective_batch_size / step_time
                 logging.info(train_tracker)
                 if wandb_logger:
+                    # Policy sub-losses (latent_loss, action_loss, ...) are aggregated into the
+                    # tracker by update_policy, so to_dict() already carries their windowed,
+                    # rank-reduced averages — no per-step output_dict passthrough needed.
                     wandb_log_dict = train_tracker.to_dict()
-                    if output_dict:
-                        wandb_log_dict.update(output_dict)
                     # Log sample weighting statistics if enabled
                     if sample_weighter is not None:
                         weighter_stats = sample_weighter.get_stats()

src/lerobot/utils/device_utils.py

@@ -59,6 +59,20 @@ def get_safe_torch_device(try_device: str, log: bool = False) -> torch.device:
     return device
 
 
+def resolve_safetensors_device(map_location: str | torch.device) -> str:
+    """Resolve a device string for a safetensors load, working around a device-mapping quirk.
+
+    safetensors' load maps the bare string "cuda" to cuda:0 regardless of the current device
+    (unlike torch's .to("cuda"), which honors torch.cuda.current_device()). Under multi-GPU
+    accelerate/FSDP every rank would then load its weights onto GPU 0, OOMing it before sharding.
+    Resolve "cuda" to the concrete current-device index so each rank loads onto its own GPU.
+    """
+    map_location = str(map_location)
+    if map_location == "cuda" and torch.cuda.is_available():
+        return f"cuda:{torch.cuda.current_device()}"
+    return map_location
+
+
 def get_safe_dtype(dtype: torch.dtype, device: str | torch.device):
     """
     mps is currently not compatible with float64

src/lerobot/utils/logging_utils.py

@@ -104,6 +104,7 @@ class MetricsTracker:
         "episodes",
         "epochs",
         "accelerator",
+        "_caller_metrics",
     ]
 
     def __init__(
@@ -129,6 +130,9 @@ class MetricsTracker:
         self.episodes = self.samples / self._avg_samples_per_ep
         self.epochs = self.samples / self._num_frames
         self.accelerator = accelerator
+        # Meter names the caller registered up front. update_metrics() leaves these untouched, so a
+        # policy that echoes e.g. "loss" in its output dict can't clobber the aggregated meter.
+        self._caller_metrics: set[str] = set(self.metrics)
 
     def __getattr__(self, name: str) -> int | dict[str, AverageMeter] | AverageMeter | Any:
         if name in self.__dict__:
@@ -156,6 +160,21 @@ class MetricsTracker:
         self.episodes = self.samples / self._avg_samples_per_ep
         self.epochs = self.samples / self._num_frames
 
+    def update_metrics(self, values: dict[str, Any]) -> None:
+        """Accumulate a dict of scalar metrics, auto-registering a meter for each new key.
+
+        Non-numeric values and bools are ignored.
+        Caller-registered metrics (those passed to the constructor) are never overridden.
+        """
+        for name, value in values.items():
+            if isinstance(value, bool) or not isinstance(value, (int, float)):
+                continue
+            if name in self._caller_metrics:
+                continue
+            if name not in self.metrics:
+                self.metrics[name] = AverageMeter(name, ":.3f", reduction="mean")
+            self.metrics[name].update(float(value))
+
     def reduce_across_ranks(self) -> None:
         """
         Synchronises the running averages of every metric whose ``reduction`` is not ``"none"``

tests/annotations/test_modules.py

@@ -85,7 +85,7 @@ def _spy_responder(captured: list[list[dict[str, Any]]], reply: Any):
 def test_module1_plan_memory_subtask_smoke(fixture_dataset_root: Path, tmp_path: Path) -> None:
     vlm = make_canned_responder(
         {
-            "atomic subtasks": {
+            "COMPLETED manipulation events": {
                 "subtasks": [
                     {"text": "grasp the handle of the sponge", "start": 0.0, "end": 0.4},
                     {"text": "wipe the counter from left to right", "start": 0.4, "end": 0.8},
@@ -126,7 +126,7 @@ def test_module1_emit_memory_false_skips_memory_keeps_subtasks_and_plan(
     leaving subtask + plan generation intact — symmetric to ``emit_plan``."""
     vlm = make_canned_responder(
         {
-            "atomic subtasks": {
+            "COMPLETED manipulation events": {
                 "subtasks": [
                     {"text": "grasp the handle of the sponge", "start": 0.0, "end": 0.4},
                     {"text": "wipe the counter from left to right", "start": 0.4, "end": 0.8},
@@ -318,7 +318,7 @@ def test_module1_attaches_contact_sheets_to_subtask_prompt(
                     return block.get("text", "")
         return ""
 
-    subtask_calls = [m for m in captured if "atomic subtasks" in _prompt_text(m)]
+    subtask_calls = [m for m in captured if "COMPLETED manipulation events" in _prompt_text(m)]
     assert len(subtask_calls) == 1, "expected exactly one subtask-prompt VLM call"
     content = subtask_calls[0][0]["content"]
     video_blocks = [b for b in content if isinstance(b, dict) and b.get("type") == "video"]

tests/scripts/test_lerobot_annotate.py

@@ -18,6 +18,8 @@ import json
 from types import SimpleNamespace
 
 import pytest
+import requests
+from huggingface_hub.errors import RevisionNotFoundError
 
 # ``lerobot.scripts.lerobot_annotate`` (and the ``_push_to_hub`` path it
 # exercises) imports ``lerobot.datasets``, which only ships under the
@@ -26,11 +28,13 @@ pytest.importorskip("datasets", reason="datasets is required (install lerobot[da
 
 
 def test_push_to_hub_tags_uploaded_dataset_revision(tmp_path, monkeypatch):
-    from lerobot.scripts.lerobot_annotate import _push_to_hub
+    from lerobot.scripts import lerobot_annotate
 
     root = tmp_path / "dataset"
     (root / "meta").mkdir(parents=True)
-    (root / "meta" / "info.json").write_text(json.dumps({"codebase_version": "v3.0"}))
+    (root / "meta" / "info.json").write_text(
+        json.dumps({"codebase_version": "v3.0", "fps": 30, "features": {}})
+    )
 
     calls = {}
 
@@ -43,9 +47,6 @@ def test_push_to_hub_tags_uploaded_dataset_revision(tmp_path, monkeypatch):
             return SimpleNamespace(oid="abc123")
 
         def delete_tag(self, repo_id, **kwargs):
-            import requests
-            from huggingface_hub.errors import RevisionNotFoundError
-
             calls["delete_tag"] = {"repo_id": repo_id, **kwargs}
             # Simulate the common case: no stale tag to delete.
             raise RevisionNotFoundError("no such tag", response=requests.Response())
@@ -53,7 +54,12 @@ def test_push_to_hub_tags_uploaded_dataset_revision(tmp_path, monkeypatch):
         def create_tag(self, **kwargs):
             calls["create_tag"] = kwargs
 
-    monkeypatch.setattr("huggingface_hub.HfApi", FakeHfApi)
+    monkeypatch.setattr(lerobot_annotate, "HfApi", FakeHfApi)
+
+    def fake_card_push(self, **kwargs):
+        calls["card_push"] = {"content": str(self), **kwargs}
+
+    monkeypatch.setattr("huggingface_hub.DatasetCard.push_to_hub", fake_card_push)
 
     cfg = SimpleNamespace(
         repo_id="source/dataset",
@@ -62,7 +68,7 @@ def test_push_to_hub_tags_uploaded_dataset_revision(tmp_path, monkeypatch):
         push_commit_message=None,
     )
 
-    _push_to_hub(root, cfg)
+    lerobot_annotate._push_to_hub(root, cfg)
 
     assert calls["create_repo"] == {
         "repo_id": "annotated/dataset",
@@ -71,6 +77,13 @@ def test_push_to_hub_tags_uploaded_dataset_revision(tmp_path, monkeypatch):
         "exist_ok": True,
     }
     assert calls["upload_folder"]["repo_id"] == "annotated/dataset"
+    # The source README must not be copied over: its links (e.g. the
+    # visualize badge) point at the source dataset. A card regenerated for
+    # the target repo is pushed instead.
+    assert "README.md" in calls["upload_folder"]["ignore_patterns"]
+    assert calls["card_push"]["repo_id"] == "annotated/dataset"
+    assert "visualize_dataset?path=annotated/dataset" in calls["card_push"]["content"]
+    assert "source/dataset" not in calls["card_push"]["content"]
     # A stale tag (e.g. from a previous annotation run) is deleted first so
     # the new tag always points at the upload we just made.
     assert calls["delete_tag"] == {

tests/utils/test_logging_utils.py

@@ -233,3 +233,37 @@ def test_metrics_tracker_reduce_across_ranks_invokes_reduce():
     # accumulate against the cluster view rather than the stale per-rank sum.
     meter = tracker.update_s
     assert meter.sum / meter.count == pytest.approx(meter.avg)
+
+
+def test_metrics_tracker_update_metrics_registers_and_averages():
+    tracker = MetricsTracker(batch_size=32, num_frames=1000, num_episodes=50, metrics={})
+    tracker.update_metrics({"latent_loss": 0.2, "action_loss": 0.4})
+    tracker.update_metrics({"latent_loss": 0.4, "action_loss": 0.6})
+
+    # New keys are auto-registered as mean-reduced meters and averaged over the window.
+    assert tracker.metrics["latent_loss"].reduction == "mean"
+    assert tracker.metrics["latent_loss"].avg == pytest.approx(0.3)
+    assert tracker.metrics["action_loss"].avg == pytest.approx(0.5)
+    assert tracker.to_dict()["latent_loss"] == pytest.approx(0.3)
+
+
+def test_metrics_tracker_update_metrics_skips_non_numeric():
+    tracker = MetricsTracker(batch_size=32, num_frames=1000, num_episodes=50, metrics={})
+    tracker.update_metrics({"loss": 0.5, "head_mode": "sparse", "enabled": True})
+
+    # strings and bools ignored
+    assert "loss" in tracker.metrics
+    assert "head_mode" not in tracker.metrics
+    assert "enabled" not in tracker.metrics
+
+
+def test_metrics_tracker_update_metrics_does_not_override_caller_meter():
+    # A policy that echoes "loss" in its output dict must not overwrite the caller-owned,
+    # already-aggregated loss meter.
+    metrics = {"loss": AverageMeter("loss", ":.3f", reduction="mean")}
+    tracker = MetricsTracker(batch_size=32, num_frames=1000, num_episodes=50, metrics=metrics)
+    tracker.loss = 1.0  # caller-set optimized loss
+    tracker.update_metrics({"loss": 99.0, "latent_loss": 0.2})
+
+    assert tracker.metrics["loss"].avg == pytest.approx(1.0)  # snapshot ignored
+    assert tracker.metrics["latent_loss"].avg == pytest.approx(0.2)