feat(training): add inline offline validation with train/eval split

- Add eval_split config for balanced per-task holdout
- Add eval_steps for periodic inline eval loss computation
- Add max_eval_samples to cap eval cost

.github/workflows/benchmark_tests.yml

@@ -167,9 +167,9 @@ jobs:
 
       # ── LIBERO TRAIN+EVAL SMOKE ──────────────────────────────────────────────
       # Train SmolVLA for 1 step (batch_size=1, dataset episode 0 only) then
-      # immediately runs eval inside the training loop (eval_freq=1, 1 episode).
+      # immediately runs eval inside the training loop (env_eval_freq=1, 1 episode).
       # Tests the full train→eval-within-training pipeline end-to-end.
-      - name: Run Libero train+eval smoke (1 step, eval_freq=1)
+      - name: Run Libero train+eval smoke (1 step, env_eval_freq=1)
         if: env.HF_USER_TOKEN != ''
         run: |
           docker run --name libero-train-smoke --gpus all \
@@ -196,7 +196,7 @@ jobs:
                 --output_dir=/tmp/train-smoke \
                 --steps=1 \
                 --batch_size=1 \
-                --eval_freq=1 \
+                --env_eval_freq=1 \
                 --eval.n_episodes=1 \
                 --eval.batch_size=1 \
                 --eval.use_async_envs=false \

Makefile

@@ -58,7 +58,7 @@ test-act-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=4 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_freq=2 \
@@ -96,7 +96,7 @@ test-diffusion-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=2 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_checkpoint=true \
@@ -126,7 +126,7 @@ test-tdmpc-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=2 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_checkpoint=true \
@@ -161,7 +161,7 @@ test-smolvla-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=4 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_freq=2 \

docs/source/hilserl.mdx

@@ -719,7 +719,7 @@ Example configuration for training the [reward classifier](https://huggingface.c
   "num_workers": 4,
   "steps": 5000,
   "log_freq": 10,
-  "eval_freq": 1000,
+  "env_eval_freq": 1000,
   "save_freq": 1000,
   "save_checkpoint": true,
   "seed": 2,

docs/source/libero.mdx

@@ -143,7 +143,7 @@ lerobot-train \
   --batch_size=4 \
   --eval.batch_size=1 \
   --eval.n_episodes=1 \
-  --eval_freq=1000
+  --env_eval_freq=1000
 ```
 
 ## Reproducing published results

docs/source/libero_plus.mdx

@@ -173,7 +173,7 @@ lerobot-train \
     --batch_size=4 \
     --eval.batch_size=1 \
     --eval.n_episodes=1 \
-    --eval_freq=1000
+    --env_eval_freq=1000
 ```
 
 ## Relationship to LIBERO

docs/source/metaworld.mdx

@@ -120,11 +120,11 @@ lerobot-train \
   --batch_size=4 \
   --eval.batch_size=1 \
   --eval.n_episodes=1 \
-  --eval_freq=1000
+  --env_eval_freq=1000
 ```
 
 ## Practical tips
 
 - Use the one-hot task conditioning for multi-task training (MT10/MT50 conventions) so policies have explicit task context.
 - Inspect the dataset task descriptions and the `info["is_success"]` keys when writing post-processing or logging so your success metrics line up with the benchmark.
-- Adjust `batch_size`, `steps`, and `eval_freq` to match your compute budget.
+- Adjust `batch_size`, `steps`, and `env_eval_freq` to match your compute budget.

docs/source/molmoact2.mdx

@@ -103,7 +103,7 @@ accelerate launch \
   --batch_size=32 \
   --num_workers=4 \
   --log_freq=20 \
-  --eval_freq=-1 \
+  --env_eval_freq=-1 \
   --save_checkpoint=true \
   --save_freq=2000
 ```
@@ -142,7 +142,7 @@ accelerate launch \
   --batch_size=32 \
   --num_workers=4 \
   --log_freq=20 \
-  --eval_freq=-1 \
+  --env_eval_freq=-1 \
   --save_checkpoint=true \
   --save_freq=2000
 ```

docs/source/multi_task_dit.mdx

@@ -314,7 +314,7 @@ lerobot-train \
   --steps=30000 \
   --save_freq=1000 \
   --log_freq=100 \
-  --eval_freq=1000 \
+  --env_eval_freq=1000 \
   --policy.type=multi_task_dit \
   --policy.device=cuda \
   --policy.horizon=32 \

docs/source/robocasa.mdx

@@ -166,7 +166,7 @@ lerobot-train \
   --output_dir=./outputs/smolvla_robocasa_CloseFridge \
   --steps=100000 \
   --batch_size=4 \
-  --eval_freq=5000 \
+  --env_eval_freq=5000 \
   --eval.batch_size=1 \
   --eval.n_episodes=5 \
   --save_freq=10000

docs/source/vlabench.mdx

@@ -165,7 +165,7 @@ lerobot-train \
   --output_dir=./outputs/smolvla_vlabench_primitive \
   --steps=100000 \
   --batch_size=4 \
-  --eval_freq=5000 \
+  --env_eval_freq=5000 \
   --eval.batch_size=1 \
   --eval.n_episodes=1 \
   --save_freq=10000

src/lerobot/configs/default.py

@@ -39,6 +39,8 @@ class DatasetConfig:
     # This reduces memory and speeds up DataLoader IPC. The training pipeline handles the conversion.
     return_uint8: bool = False
     streaming: bool = False
+    # Fraction of episodes held out per task for offline evaluation (0.0 = disabled).
+    eval_split: float = 0.0
 
     def __post_init__(self) -> None:
         if self.episodes is not None:
@@ -73,8 +75,6 @@ class EvalConfig:
     # `use_async_envs` specifies whether to use asynchronous environments (multiprocessing).
     # Defaults to True; automatically downgraded to SyncVectorEnv when batch_size=1.
     use_async_envs: bool = True
-    # Whether to record eval rollouts as a LeRobot v3.0 dataset on disk.
-    recording: bool = False
 
     def __post_init__(self) -> None:
         if self.batch_size == 0:

src/lerobot/configs/train.py

@@ -100,8 +100,13 @@ class TrainPipelineConfig(HubMixin):
     prefetch_factor: int = 4
     persistent_workers: bool = True
     steps: int = 100_000
-    eval_freq: int = 20_000
+    # Run policy in the simulation environment every N steps to measure reward/success (0 = disabled).
+    env_eval_freq: int = 20_000
     log_freq: int = 200
+    # Compute eval loss on held-out episodes every N steps (0 = disabled). Requires eval_split > 0.
+    eval_steps: int = 0
+    # Cap on total eval samples, split uniformly across tasks (0 = use all held-out data).
+    max_eval_samples: int = 0
     tolerance_s: float = 1e-4
     save_checkpoint: bool = True
     # Checkpoint is saved every `save_freq` training iterations and after the last training step.

src/lerobot/datasets/__init__.py

@@ -35,7 +35,7 @@ from .dataset_tools import (
     remove_feature,
     split_dataset,
 )
-from .factory import make_dataset, resolve_delta_timestamps
+from .factory import make_dataset, make_train_eval_datasets, resolve_delta_timestamps
 from .image_writer import safe_stop_image_writer
 from .io_utils import load_episodes, write_stats
 from .language import (
@@ -89,6 +89,7 @@ __all__ = [
     "get_feature_stats",
     "load_episodes",
     "make_dataset",
+    "make_train_eval_datasets",
     "merge_datasets",
     "modify_features",
     "modify_tasks",

src/lerobot/datasets/factory.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import logging
+import math
 from pprint import pformat
 
 import torch
@@ -130,3 +131,81 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
                 dataset.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
 
     return dataset
+
+
+def make_train_eval_datasets(
+    cfg: TrainPipelineConfig,
+) -> tuple[LeRobotDataset | MultiLeRobotDataset, LeRobotDataset | None]:
+    """Create train and optional eval datasets by splitting episodes based on eval_split.
+
+    The last ceil(n_episodes * eval_split) episodes per task are held out for evaluation.
+    If eval_split == 0.0, returns (full_dataset, None).
+    """
+    full_dataset = make_dataset(cfg)
+
+    if cfg.dataset.eval_split == 0.0:
+        return full_dataset, None
+
+    base_episodes = (
+        full_dataset.episodes if full_dataset.episodes is not None else list(range(full_dataset.num_episodes))
+    )
+
+    episode_tasks = full_dataset.meta.episodes["tasks"]
+    task_to_episodes: dict[str, list[int]] = {}
+    for ep_idx in base_episodes:
+        task_key = episode_tasks[ep_idx][0] if episode_tasks[ep_idx] else ""
+        task_to_episodes.setdefault(task_key, []).append(ep_idx)
+
+    train_episodes, eval_episodes = [], []
+    for eps in task_to_episodes.values():
+        n_eval = math.ceil(len(eps) * cfg.dataset.eval_split)
+        train_episodes.extend(eps[: len(eps) - n_eval])
+        eval_episodes.extend(eps[len(eps) - n_eval :])
+
+    if not train_episodes:
+        raise ValueError(
+            f"eval_split={cfg.dataset.eval_split} leaves 0 training episodes from {len(base_episodes)} total."
+        )
+
+    logging.info(
+        f"Train/eval split: {len(train_episodes)} train, {len(eval_episodes)} eval "
+        f"(eval_split={cfg.dataset.eval_split}, {len(task_to_episodes)} tasks)"
+    )
+
+    delta_timestamps = resolve_delta_timestamps(cfg.trainable_config, full_dataset.meta)
+
+    train_image_transforms = (
+        ImageTransforms(cfg.dataset.image_transforms) if cfg.dataset.image_transforms.enable else None
+    )
+
+    train_dataset = LeRobotDataset(
+        cfg.dataset.repo_id,
+        root=cfg.dataset.root,
+        episodes=train_episodes,
+        delta_timestamps=delta_timestamps,
+        image_transforms=train_image_transforms,
+        revision=cfg.dataset.revision,
+        video_backend=cfg.dataset.video_backend,
+        return_uint8=True,
+        tolerance_s=cfg.tolerance_s,
+    )
+
+    eval_dataset = LeRobotDataset(
+        cfg.dataset.repo_id,
+        root=cfg.dataset.root,
+        episodes=eval_episodes,
+        delta_timestamps=delta_timestamps,
+        image_transforms=None,
+        revision=cfg.dataset.revision,
+        video_backend=cfg.dataset.video_backend,
+        return_uint8=True,
+        tolerance_s=cfg.tolerance_s,
+    )
+
+    if cfg.dataset.use_imagenet_stats:
+        for ds in (train_dataset, eval_dataset):
+            for key in ds.meta.camera_keys:
+                for stats_type, stats in IMAGENET_STATS.items():
+                    ds.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
+
+    return train_dataset, eval_dataset

src/lerobot/scripts/lerobot_eval.py

@@ -74,7 +74,6 @@ from tqdm import trange
 
 from lerobot.configs import parser
 from lerobot.configs.eval import EvalPipelineConfig
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.envs import (
     check_env_attributes_and_types,
     close_envs,
@@ -85,7 +84,7 @@ from lerobot.envs import (
 from lerobot.policies import PreTrainedPolicy, make_policy, make_pre_post_processors
 from lerobot.processor import PolicyProcessorPipeline
 from lerobot.types import PolicyAction
-from lerobot.utils.constants import ACTION, DONE, OBS_IMAGE, OBS_IMAGES, OBS_STATE, OBS_STR, REWARD
+from lerobot.utils.constants import ACTION, DONE, OBS_STR, REWARD
 from lerobot.utils.device_utils import get_safe_torch_device
 from lerobot.utils.import_utils import register_third_party_plugins
 from lerobot.utils.io_utils import write_video
@@ -96,56 +95,6 @@ from lerobot.utils.utils import (
 )
 
 
-def _env_features_to_dataset_features(env_features: dict) -> dict:
-    """Convert EnvConfig.features (PolicyFeature objects) to the plain dict format for LeRobotDataset.create()."""
-    features = {}
-    for key, ft in env_features.items():
-        if ft.type.value == "visual":
-            features[key] = {
-                "dtype": "video",
-                "shape": tuple(ft.shape),
-                "names": ["channel", "height", "width"],
-            }
-        else:
-            features[key] = {"dtype": "float32", "shape": tuple(ft.shape), "names": None}
-    features["next.reward"] = {"dtype": "float32", "shape": (1,), "names": None}
-    features["next.success"] = {"dtype": "bool", "shape": (1,), "names": None}
-    features["next.done"] = {"dtype": "bool", "shape": (1,), "names": None}
-    return features
-
-
-def _build_raw_frame(
-    raw_obs: dict,
-    env_idx: int,
-    action: np.ndarray,
-    reward: float,
-    success: bool,
-    done: bool,
-    task: str,
-) -> dict:
-    """Build a dataset frame from raw env observations for one env index."""
-    frame: dict[str, Any] = {}
-    if "pixels" in raw_obs:
-        if isinstance(raw_obs["pixels"], dict):
-            for cam_name, img in raw_obs["pixels"].items():
-                frame[f"{OBS_IMAGES}.{cam_name}"] = img[env_idx]
-        else:
-            frame[OBS_IMAGE] = raw_obs["pixels"][env_idx]
-    if "agent_pos" in raw_obs:
-        frame[OBS_STATE] = raw_obs["agent_pos"][env_idx]
-    for key, val in raw_obs.items():
-        if key in ("pixels", "agent_pos"):
-            continue
-        if isinstance(val, np.ndarray):
-            frame[f"{OBS_STR}.{key}"] = val[env_idx]
-    frame[ACTION] = action
-    frame["next.reward"] = np.float32(reward)
-    frame["next.success"] = success
-    frame["next.done"] = done
-    frame["task"] = task
-    return frame
-
-
 def rollout(
     env: gym.vector.VectorEnv,
     policy: PreTrainedPolicy,
@@ -156,7 +105,6 @@ def rollout(
     seeds: list[int] | None = None,
     return_observations: bool = False,
     render_callback: Callable[[gym.vector.VectorEnv], None] | None = None,
-    recording_dataset: Any | None = None,
 ) -> dict:
     """Run a batched policy rollout once through a batch of environments.
 
@@ -197,14 +145,6 @@ def rollout(
     if render_callback is not None:
         render_callback(env)
 
-    raw_observation = deepcopy(observation) if recording_dataset is not None else None
-    task_desc = ""
-    if recording_dataset is not None:
-        try:
-            task_desc = list(env.call("task_description"))[0]
-        except (AttributeError, NotImplementedError):
-            task_desc = ""
-
     all_observations = []
     all_actions = []
     all_rewards = []
@@ -277,25 +217,6 @@ def rollout(
         else:
             successes = [False] * env.num_envs
 
-        if recording_dataset is not None and raw_observation is not None:
-            prev_done = done.copy()
-            for env_idx in range(env.num_envs):
-                if prev_done[env_idx]:
-                    continue
-                frame = _build_raw_frame(
-                    raw_observation,
-                    env_idx,
-                    action_numpy[env_idx],
-                    reward[env_idx],
-                    successes[env_idx],
-                    bool(terminated[env_idx] | truncated[env_idx]),
-                    task_desc,
-                )
-                recording_dataset.add_frame(frame)
-                if terminated[env_idx] or truncated[env_idx]:
-                    recording_dataset.save_episode()
-            raw_observation = deepcopy(observation)
-
         # Keep track of which environments are done so far.
         # Mark the episode as done if we reach the maximum step limit.
         # This ensures that the rollout always terminates cleanly at `max_steps`,
@@ -352,7 +273,6 @@ def eval_policy(
     videos_dir: Path | None = None,
     return_episode_data: bool = False,
     start_seed: int | None = None,
-    recording_dataset: Any | None = None,
 ) -> dict:
     """
     Args:
@@ -441,7 +361,6 @@ def eval_policy(
             seeds=list(seeds) if seeds else None,
             return_observations=return_episode_data,
             render_callback=render_frame if max_episodes_rendered > 0 else None,
-            recording_dataset=recording_dataset,
         )
 
         # Figure out where in each rollout sequence the first done condition was encountered (results after
@@ -644,10 +563,6 @@ def eval_main(cfg: EvalPipelineConfig):
     # Create environment-specific preprocessor and postprocessor (e.g., for LIBERO environments)
     env_preprocessor, env_postprocessor = make_env_pre_post_processors(env_cfg=cfg.env, policy_cfg=cfg.policy)
 
-    recording_dir = Path(cfg.output_dir) / "recordings" if cfg.eval.recording else None
-    max_episodes_rendered = 0 if cfg.eval.recording else 10
-    videos_dir = None if cfg.eval.recording else Path(cfg.output_dir) / "videos"
-
     with torch.no_grad(), torch.autocast(device_type=device.type) if cfg.policy.use_amp else nullcontext():
         info = eval_policy_all(
             envs=envs,
@@ -657,13 +572,10 @@ def eval_main(cfg: EvalPipelineConfig):
             preprocessor=preprocessor,
             postprocessor=postprocessor,
             n_episodes=cfg.eval.n_episodes,
-            max_episodes_rendered=max_episodes_rendered,
-            videos_dir=videos_dir,
-            return_episode_data=False,
+            max_episodes_rendered=10,
+            videos_dir=Path(cfg.output_dir) / "videos",
             start_seed=cfg.seed,
             max_parallel_tasks=cfg.env.max_parallel_tasks,
-            recording_dir=recording_dir,
-            env_features=cfg.env.features if cfg.eval.recording else None,
         )
         print("Overall Aggregated Metrics:")
         print(info["overall"])
@@ -706,7 +618,6 @@ def eval_one(
     videos_dir: Path | None,
     return_episode_data: bool,
     start_seed: int | None,
-    recording_dataset: Any | None = None,
 ) -> TaskMetrics:
     """Evaluates one task_id of one suite using the provided vec env."""
 
@@ -724,7 +635,6 @@ def eval_one(
         videos_dir=task_videos_dir,
         return_episode_data=return_episode_data,
         start_seed=start_seed,
-        recording_dataset=recording_dataset,
     )
 
     per_episode = task_result["per_episode"]
@@ -751,8 +661,6 @@ def run_one(
     videos_dir: Path | None,
     return_episode_data: bool,
     start_seed: int | None,
-    recording_dir: Path | None = None,
-    env_features: dict | None = None,
 ):
     """
     Run eval_one for a single (task_group, task_id, env).
@@ -764,38 +672,21 @@ def run_one(
         task_videos_dir = videos_dir / f"{task_group}_{task_id}"
         task_videos_dir.mkdir(parents=True, exist_ok=True)
 
-    recording_dataset = None
-    if recording_dir is not None and env_features is not None:
-        task_recording_dir = recording_dir / f"{task_group}_{task_id}"
-        fps = env.unwrapped.metadata.get("render_fps", 30)
-        features = _env_features_to_dataset_features(env_features)
-        recording_dataset = LeRobotDataset.create(
-            repo_id=f"eval_{task_group}_{task_id}",
-            fps=fps,
-            features=features,
-            root=str(task_recording_dir),
-            use_videos=True,
-        )
-
-    try:
-        metrics = eval_one(
-            env,
-            policy=policy,
-            env_preprocessor=env_preprocessor,
-            env_postprocessor=env_postprocessor,
-            preprocessor=preprocessor,
-            postprocessor=postprocessor,
-            n_episodes=n_episodes,
-            max_episodes_rendered=max_episodes_rendered,
-            videos_dir=task_videos_dir,
-            return_episode_data=return_episode_data,
-            start_seed=start_seed,
-            recording_dataset=recording_dataset,
-        )
-    finally:
-        if recording_dataset is not None:
-            recording_dataset.finalize()
-
+    # Call the existing eval_one (assumed to return TaskMetrics-like dict)
+    metrics = eval_one(
+        env,
+        policy=policy,
+        env_preprocessor=env_preprocessor,
+        env_postprocessor=env_postprocessor,
+        preprocessor=preprocessor,
+        postprocessor=postprocessor,
+        n_episodes=n_episodes,
+        max_episodes_rendered=max_episodes_rendered,
+        videos_dir=task_videos_dir,
+        return_episode_data=return_episode_data,
+        start_seed=start_seed,
+    )
+    # ensure we always provide video_paths key to simplify accumulation
     if max_episodes_rendered > 0:
         metrics.setdefault("video_paths", [])
     return task_group, task_id, metrics
@@ -811,8 +702,6 @@ def eval_policy_all(
     n_episodes: int,
     *,
     max_episodes_rendered: int = 0,
-    recording_dir: Path | None = None,
-    env_features: dict | None = None,
     videos_dir: Path | None = None,
     return_episode_data: bool = False,
     start_seed: int | None = None,
@@ -872,8 +761,6 @@ def eval_policy_all(
         videos_dir=videos_dir,
         return_episode_data=return_episode_data,
         start_seed=start_seed,
-        recording_dir=recording_dir,
-        env_features=env_features,
     )
 
     if max_parallel_tasks <= 1:

src/lerobot/scripts/lerobot_train.py

@@ -45,7 +45,8 @@ from lerobot.common.train_utils import (
 from lerobot.common.wandb_utils import WandBLogger
 from lerobot.configs import parser
 from lerobot.configs.train import TrainPipelineConfig
-from lerobot.datasets import EpisodeAwareSampler, compute_sampler_state, make_dataset
+from lerobot.datasets import EpisodeAwareSampler, compute_sampler_state
+from lerobot.datasets.factory import make_train_eval_datasets
 from lerobot.envs import close_envs, make_env, make_env_pre_post_processors
 from lerobot.optim.factory import make_optimizer_and_scheduler
 from lerobot.policies import PreTrainedPolicy, make_policy, make_pre_post_processors
@@ -244,19 +245,19 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     # LeRobotDataset skips its snapshot_download when try_load() succeeds, so no rank re-downloads.
     if is_main_process:
         logging.info("Creating dataset")
-        dataset = make_dataset(cfg)
+        dataset, eval_dataset = make_train_eval_datasets(cfg)
 
     accelerator.wait_for_everyone()
 
     # Other ranks read from the shared copy populated by the main process.
     if not is_main_process:
-        dataset = make_dataset(cfg)
+        dataset, eval_dataset = make_train_eval_datasets(cfg)
 
     # Create environment used for evaluating checkpoints during training on simulation data.
     # On real-world data, no need to create an environment as evaluations are done outside train.py,
     # using the eval.py instead, with gym_dora environment and dora-rs.
     eval_env = None
-    if cfg.eval_freq > 0 and cfg.env is not None and is_main_process:
+    if cfg.env_eval_freq > 0 and cfg.env is not None and is_main_process:
         logging.info("Creating env")
         eval_env = make_env(cfg.env, n_envs=cfg.eval.batch_size, use_async_envs=cfg.eval.use_async_envs)
 
@@ -434,6 +435,8 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
                     f"Resuming data order at epoch {sampler_state['epoch']}, "
                     f"sample {sampler_state['start_index']}"
                 )
+        if dataset.reader._absolute_to_relative_idx is not None:
+            sampler.indices = [dataset.reader._absolute_to_relative_idx[i] for i in sampler.indices]
     else:
         shuffle = True
         sampler = None
@@ -455,6 +458,31 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         persistent_workers=cfg.persistent_workers and cfg.num_workers > 0,
     )
 
+    # Build eval dataloader if a held-out split exists
+    eval_dataloader = None
+    if eval_dataset is not None:
+        eval_ds = eval_dataset
+        if cfg.max_eval_samples > 0 and hasattr(eval_dataset, "hf_dataset"):
+            task_indices = eval_dataset.hf_dataset["task_index"]
+            unique_tasks = sorted(set(task_indices))
+            per_task = max(1, cfg.max_eval_samples // len(unique_tasks))
+            selected: list[int] = []
+            for t in unique_tasks:
+                frames = [i for i, ti in enumerate(task_indices) if ti == t][:per_task]
+                selected.extend(frames)
+            eval_ds = torch.utils.data.Subset(eval_dataset, selected)
+
+        eval_collate_fn = lerobot_collate_fn if dataset.meta.has_language_columns else None
+        eval_dataloader = torch.utils.data.DataLoader(
+            eval_ds,
+            batch_size=cfg.batch_size,
+            shuffle=False,
+            num_workers=cfg.num_workers,
+            pin_memory=device.type == "cuda",
+            drop_last=False,
+            collate_fn=eval_collate_fn,
+        )
+
     # Prepare everything with accelerator
     accelerator.wait_for_everyone()
     policy, optimizer, dataloader, lr_scheduler = accelerator.prepare(
@@ -534,7 +562,8 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         train_tracker.step()
         is_log_step = cfg.log_freq > 0 and step % cfg.log_freq == 0
         is_saving_step = step % cfg.save_freq == 0 or step == cfg.steps
-        is_eval_step = cfg.eval_freq > 0 and step % cfg.eval_freq == 0
+        is_env_eval_step = cfg.env_eval_freq > 0 and step % cfg.env_eval_freq == 0
+        is_eval_step = cfg.eval_steps > 0 and eval_dataloader is not None and step % cfg.eval_steps == 0
 
         if is_log_step:
             # Collective reduce must run on every rank, before the main-process gate below.
@@ -557,6 +586,27 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
                     wandb_logger.log_dict(wandb_log_dict, step)
             train_tracker.reset_averages()
 
+        if is_eval_step:
+            policy.eval()
+            eval_loss_sum = 0.0
+            n_eval_batches = 0
+            with torch.no_grad(), accelerator.autocast():
+                for eval_batch in eval_dataloader:
+                    for cam_key in dataset.meta.camera_keys:
+                        if cam_key in eval_batch and eval_batch[cam_key].dtype == torch.uint8:
+                            eval_batch[cam_key] = eval_batch[cam_key].to(dtype=torch.float32) / 255.0
+                    eval_batch = preprocessor(eval_batch)
+                    loss, _ = policy.forward(eval_batch)
+                    eval_loss_sum += loss.item()
+                    n_eval_batches += 1
+            eval_loss = eval_loss_sum / max(n_eval_batches, 1)
+            policy.train()
+
+            if is_main_process:
+                logging.info(f"step {step}: eval_loss={eval_loss:.4f}")
+                if wandb_logger:
+                    wandb_logger.log_dict({"eval_loss": eval_loss}, step=step, mode="eval")
+
         if cfg.save_checkpoint and is_saving_step:
             if is_main_process:
                 logging.info(f"Checkpoint policy after step {step}")
@@ -579,7 +629,7 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
 
             accelerator.wait_for_everyone()
 
-        if cfg.env and is_eval_step:
+        if cfg.env and is_env_eval_step:
             if is_main_process:
                 step_id = get_step_identifier(step, cfg.steps)
                 logging.info(f"Eval policy at step {step}")

tests/training/test_multi_gpu.py

@@ -134,7 +134,7 @@ class TestMultiGPUTraining:
                 f"--output_dir={output_dir}",
                 "--batch_size=4",
                 "--steps=10",
-                "--eval_freq=-1",
+                "--env_eval_freq=-1",
                 "--log_freq=5",
                 "--save_freq=10",
                 "--seed=42",
@@ -177,7 +177,7 @@ class TestMultiGPUTraining:
                 f"--output_dir={output_dir}",
                 "--batch_size=4",
                 "--steps=20",
-                "--eval_freq=-1",
+                "--env_eval_freq=-1",
                 "--log_freq=5",
                 "--save_freq=10",
                 "--seed=42",