chore(training): align smolvla2_hirobot.slurm with what's actually run

Match the operator's current training command for the _tool6 retrain:

  * default DATASET / POLICY_REPO_ID / JOB_NAME point at the tool6
    iteration (super_poulain_full_tool3 → smolvla2_hirobot_super_poulain_tool6)
  * STEPS default 2000 (short enough to iterate; bump to 10k for full)
  * save_freq=$STEPS so the only checkpoint is the final one
  * OUTPUT_DIR includes step count so successive runs don't clobber
  * Drop the wider augmentation envelope I added earlier — back to
    default ColorJitter ranges (brightness ±20% etc) since the
    high_level_subtask recipe fix (current-subtask supervision) is
    expected to fix the LM-head collapse on its own; the augmentation
    is just the standard regulariser, not a load-bearing widener.
  * prompt-dropout fractions stay at the original 0.15 / 0.15 / 0.20.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

examples/training/smolvla2_hirobot.slurm

@@ -6,27 +6,23 @@
 #SBATCH --ntasks=1
 #SBATCH --gpus-per-task=8
 
-# SmolVLA2 training on an annotated dataset, with image augmentation
+# SmolVLA2 training on an annotated dataset.
-# and per-component prompt dropout enabled — the two regularisers
-# that move the model away from the "text_loss=6e-6 memorised one
-# epoch worth of frames" failure mode toward "learns concepts, not
-# pixels".
 #
-# What the regularisers do:
+# The high_level_subtask recipe (recipes/smolvla2_hirobot.yaml) was
+# fixed in PR3 to supervise the LM head with the *current* active
+# subtask span at every frame, not the next-span target which is
+# empty on stable phases. With the old recipe the head learned to
+# emit ``\n`` on every chunk boundary; the new one supervises a
+# real, scene-grounded string at every frame.
 #
-#   * --dataset.image_transforms.enable=true: applies torchvision
+# Two regularisers are still on:
-#     v2 ColorJitter (brightness/contrast/saturation/hue),
-#     SharpnessJitter and RandomAffine per frame at training time.
-#     Set max_num_transforms to control how many are sampled per
-#     frame; defaults to 3 of the 6.
-#   * --policy.plan_dropout_prob / memory / subtask: at training,
-#     randomly drop the context messages that carry the named
-#     binding so the model is forced to handle missing/stale context.
-#     Mirrors Pi0.7's prompt-component dropout (§V.E).
 #
-# Expected effect: text_loss plateaus higher (~0.5-2.0 instead of
+#   * --dataset.image_transforms.enable=true: torchvision-v2
-# ~1e-5) and the model handles slight prompt/scene drift at
+#     ColorJitter + SharpnessJitter + RandomAffine per frame; default
-# inference instead of collapsing to memorised fragments.
+#     envelope (brightness ±20% etc).
+#   * --policy.{plan,memory,subtask}_dropout_prob: randomly drop the
+#     context messages carrying the named recipe binding so the model
+#     handles missing/stale context. Mirrors Pi0.7 §V.E.
 
 set -euo pipefail
 
@@ -39,20 +35,20 @@ export HF_HUB_DOWNLOAD_TIMEOUT="${HF_HUB_DOWNLOAD_TIMEOUT:-120}"
 export WANDB_INIT_TIMEOUT="${WANDB_INIT_TIMEOUT:-300}"
 
 DATASET="${DATASET:-pepijn223/super_poulain_full_tool3}"
-POLICY_REPO_ID="${POLICY_REPO_ID:-pepijn223/smolvla2_hirobot_super_poulain_tool4}"
+POLICY_REPO_ID="${POLICY_REPO_ID:-pepijn223/smolvla2_hirobot_super_poulain_tool6}"
-JOB_NAME="${JOB_NAME:-smolvla2-hirobot-super-poulain-tool4}"
+JOB_NAME="${JOB_NAME:-smolvla2-hirobot-super-poulain-tool6}"
 NUM_PROCESSES="${NUM_PROCESSES:-8}"
 BATCH_SIZE="${BATCH_SIZE:-32}"
-STEPS="${STEPS:-10000}"
+STEPS="${STEPS:-2000}"
 RUN_ID="${SLURM_JOB_ID:-$(date +%Y%m%d_%H%M%S)}"
-OUTPUT_DIR="${OUTPUT_DIR:-/fsx/pepijn/outputs/train/smolvla2_hirobot_${RUN_ID}}"
+OUTPUT_DIR="${OUTPUT_DIR:-/fsx/pepijn/outputs/train/smolvla2_hirobot_super_poulain_tool3_${STEPS}_${RUN_ID}}"
 
 echo "Training smolvla2 on $DATASET"
 echo "  GPUs:         $NUM_PROCESSES"
 echo "  batch:        $BATCH_SIZE / GPU (global=$((NUM_PROCESSES * BATCH_SIZE)))"
 echo "  steps:        $STEPS"
 echo "  output:       $OUTPUT_DIR"
-echo "  augmentation: image_transforms ON (wide), prompt dropout {plan:0.20 memory:0.20 subtask:0.30}"
+echo "  augmentation: image_transforms ON, prompt dropout {plan:0.15 memory:0.15 subtask:0.20}"
 
 accelerate launch --multi_gpu --num_processes="$NUM_PROCESSES" \
     -m lerobot.scripts.lerobot_train \
@@ -61,17 +57,6 @@ accelerate launch --multi_gpu --num_processes="$NUM_PROCESSES" \
     --dataset.repo_id="$DATASET" \
     --dataset.revision=main \
     --dataset.video_backend=pyav \
-    --dataset.image_transforms.enable=true \
-    --dataset.image_transforms.max_num_transforms=4 \
-    --dataset.image_transforms.random_order=true \
-    --dataset.image_transforms.tfs.brightness.kwargs='{"brightness": [0.5, 1.6]}' \
-    --dataset.image_transforms.tfs.contrast.kwargs='{"contrast": [0.6, 1.5]}' \
-    --dataset.image_transforms.tfs.saturation.kwargs='{"saturation": [0.3, 1.7]}' \
-    --dataset.image_transforms.tfs.hue.kwargs='{"hue": [-0.1, 0.1]}' \
-    --dataset.image_transforms.tfs.affine.kwargs='{"degrees": [-15.0, 15.0], "translate": [0.15, 0.15]}' \
-    --policy.plan_dropout_prob=0.20 \
-    --policy.memory_dropout_prob=0.20 \
-    --policy.subtask_dropout_prob=0.30 \
     --output_dir="$OUTPUT_DIR" \
     --job_name="$JOB_NAME" \
     --policy.repo_id="$POLICY_REPO_ID" \
@@ -85,5 +70,11 @@ accelerate launch --multi_gpu --num_processes="$NUM_PROCESSES" \
     --wandb.disable_artifact=true \
     --wandb.project=hirobot \
     --log_freq=100 \
-    --save_freq=1000 \
+    --save_freq="$STEPS" \
-    --num_workers=0
+    --num_workers=0 \
+    --dataset.image_transforms.enable=true \
+    --dataset.image_transforms.max_num_transforms=3 \
+    --dataset.image_transforms.random_order=true \
+    --policy.plan_dropout_prob=0.15 \
+    --policy.memory_dropout_prob=0.15 \
+    --policy.subtask_dropout_prob=0.20