From c424d906cefb27ab3dfca38c7982f6d4c1531780 Mon Sep 17 00:00:00 2001 From: Khalil Meftah Date: Tue, 21 Apr 2026 14:31:40 +0200 Subject: [PATCH] fix(train): raise ValueError for non-trainable reward models in train function --- src/lerobot/scripts/lerobot_train.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/src/lerobot/scripts/lerobot_train.py b/src/lerobot/scripts/lerobot_train.py index ac3914637..be99b70bb 100644 --- a/src/lerobot/scripts/lerobot_train.py +++ b/src/lerobot/scripts/lerobot_train.py @@ -260,6 +260,11 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None): dataset_stats=dataset.meta.stats, dataset_meta=dataset.meta, ) + if not policy.is_trainable: + raise ValueError( + f"Reward model '{policy.name}' is zero-shot and cannot be trained via lerobot-train. " + "Use it directly for inference via compute_reward() (e.g. offline precompute)." + ) else: if is_main_process: logging.info("Creating policy")