Add stack bowl RECAP value checkpoint metadata

franka/rlinf/20260903-202954-franka_stack_bowl_pi05_recap_value_sft_fail300/README.md

@@ -0,0 +1,28 @@
+# Franka Stack Bowl RECAP Value SFT Checkpoint
+
+This checkpoint is a RECAP value/critic model trained for the Franka stack-bowl task using labeled rollout data.
+
+Important: this is not a direct action-policy checkpoint. It is the RECAP value model checkpoint produced by `examples/offline_rl/advantage_labeling/recap/train_value.py`.
+
+## Run
+
+- Model family in config: `pi05`
+- Task: `Stack the three bowls in size order: the purple bowl first, then the beige bowl.`
+- Dataset: stack-bowl RECAP rollout labels, 34 episodes, 196579 frames
+- Labels: 22 success / 12 failure
+- Return sidecar: `returns_fail300.parquet`
+- Training: resumed from `global_step_5000`, completed to `global_step_8000`
+- Final checkpoint: `global_step_8000`
+
+## Final eval metrics
+
+- `eval/loss`: 2.1
+- `eval/mae`: 0.0219
+- `eval/cat_acc_best`: 0.287
+- `eval/cat_acc_neighbor`: 0.44
+
+## Files
+
+- `checkpoints/global_step_8000/actor/model_state_dict/full_weights.pt`: full actor/value model state dict
+- `run_value_sft_fail300_resume5000.sh`: resume training command
+- `value_sft_fail300_resume5000_mb2_tmux.log`: training log from the successful resume segment

franka/rlinf/20260903-202954-franka_stack_bowl_pi05_recap_value_sft_fail300/checkpoints/global_step_8000/actor/model_state_dict/full_weights.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c42f0df0c54fa5003707307333cdb040c9a653335279a03aae487f1aa6653c6
+size 1799192037

franka/rlinf/20260903-202954-franka_stack_bowl_pi05_recap_value_sft_fail300/run_manifest.json

@@ -0,0 +1,23 @@
+{
+  "checkpoint_type": "recap_value_sft",
+  "policy_executable": false,
+  "model_type": "pi05",
+  "robot_type": "franka",
+  "task": "stack_bowls",
+  "prompt": "Stack the three bowls in size order: the purple bowl first, then the beige bowl.",
+  "dataset_path_on_amax": "/data2/yangky/test/recap_stack_bowls/datasets/stack_bowls_recap_rollout_rc",
+  "num_episodes": 34,
+  "num_frames": 196579,
+  "success_episodes": 22,
+  "failure_episodes": 12,
+  "returns_tag": "fail300",
+  "return_min": -20713,
+  "return_max": 0,
+  "max_steps": 8000,
+  "resume_from": "global_step_5000",
+  "final_checkpoint": "global_step_8000",
+  "eval_loss": 2.1,
+  "eval_mae": 0.0219,
+  "eval_cat_acc_best": 0.287,
+  "eval_cat_acc_neighbor": 0.44
+}

franka/rlinf/20260903-202954-franka_stack_bowl_pi05_recap_value_sft_fail300/run_value_sft_fail300_resume5000.sh

@@ -0,0 +1,37 @@
+set -euo pipefail
+source /data/lfwj/miniconda3/etc/profile.d/conda.sh
+conda activate lerobot
+cd /data/yangky/test/RealWorld-RLinf
+export RUN_ROOT=/data2/yangky/test/recap_stack_bowls
+export REPO_PATH=/data/yangky/test/RealWorld-RLinf
+export OPENPI_ARCH=/home/amax/.cache/uv/archive-v0/ZQ6tV-bbWGFB5mqO
+export PYTHONNOUSERSITE=1
+export PYTHONPATH=$OPENPI_ARCH:$REPO_PATH:/data/linjianqi/RoboTwin/policy/pi0/packages/openpi-client/src:${PYTHONPATH:-}
+export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:${LD_LIBRARY_PATH:-}
+export CUDA_VISIBLE_DEVICES=0
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+export HF_HOME=$RUN_ROOT/hf_home
+export HF_DATASETS_CACHE=$RUN_ROOT/hf_datasets_cache
+export TRANSFORMERS_CACHE=$RUN_ROOT/hf_home/transformers
+export MUJOCO_GL=egl
+export PYOPENGL_PLATFORM=egl
+export AV_LOG_FORCE_NOCOLOR=1
+export LIBAV_LOG_LEVEL=quiet
+export OPENCV_LOG_LEVEL=off
+ray stop -f || true
+mkdir -p "/data2/yangky/test/rsb_ray" "$HF_HOME" "$HF_DATASETS_CACHE" "$TRANSFORMERS_CACHE"
+ray start --head --include-dashboard=true --dashboard-host=127.0.0.1 --num-gpus=1 --temp-dir="/data2/yangky/test/rsb_ray" --disable-usage-stats
+python examples/offline_rl/advantage_labeling/recap/train_value.py --config-path /data/yangky/test/RealWorld-RLinf/examples/offline_rl/config --config-name recap_value_model_sft \
+  data.tag=fail300 +data.return_min=-20713 +data.return_max=0 \
+  data.train_data_paths='[{dataset_path: /data2/yangky/test/recap_stack_bowls/datasets/stack_bowls_recap_rollout_rc, type: rollout, weight: 1.0, robot_type: franka, model_type: pi05}]' \
+  data.eval_data_paths='[{dataset_path: /data2/yangky/test/recap_stack_bowls/datasets/stack_bowls_recap_rollout_rc, max_samples: 4096, robot_type: franka, model_type: pi05}]' \
+  data.robot_type=franka data.model_type=pi05 data.action_dim=7 data.train_num_workers=2 data.eval_num_workers=1 \
+  actor.micro_batch_size=2 actor.global_batch_size=16 actor.model.action_dim=7 \
+  actor.model.siglip_path=/data2/yangky/test/recap_stack_bowls/models/recap_value_backbones/siglip2-so400m-patch14-224 \
+  actor.model.gemma3_path=/data2/yangky/test/recap_stack_bowls/models/recap_value_backbones/gemma-3-270m \
+  actor.model.tokenizer_path=/data2/yangky/test/recap_stack_bowls/models/recap_value_backbones/gemma-3-270m \
+  runner.max_steps=8000 runner.save_interval=1000 runner.val_check_interval=500 actor.optim.total_training_steps=8000 \
+  \
+  +runner.resume_dir=/data2/yangky/test/recap_stack_bowls/logs/recap_stack_bowls_training/value_sft/stack_bowls_recap_value_sft_fail300/checkpoints/global_step_5000 \
+  runner.logger.log_path=/data2/yangky/test/recap_stack_bowls/logs/recap_stack_bowls_training/value_sft \
+  runner.logger.experiment_name=stack_bowls_recap_value_sft_fail300