Initial public release

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 BAS-VLA Contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

NOTICE.md

@@ -0,0 +1,9 @@
+# Artifact sources
+
+The residual adapter is a BAS-VLA project training artifact from the OpenVLA / LIBERO-10 offline action-calibration setting. Project contributions use the MIT License in `LICENSE`.
+
+The 400 action triplets, original agent-view observations, sampled simulator states, demonstration initial states, and two BDDL scenes are derived from [LIBERO](https://github.com/Lifelong-Robot-Learning/LIBERO) and its [demonstration dataset](https://huggingface.co/datasets/yifengzhu-hf/LIBERO-datasets). The exact source demonstration filename, demo ID and frame index are retained in each record. LIBERO's copyright and MIT notice are preserved in `licenses/LIBERO-MIT.txt`.
+
+The base carrier is [openvla/openvla-7b](https://huggingface.co/openvla/openvla-7b). The bundled normalization statistics come from [moojink/openvla-7b-oft-finetuned-libero-10](https://huggingface.co/moojink/openvla-7b-oft-finetuned-libero-10); their source is recorded in `normalization/source.json`, with the OpenVLA-OFT MIT notice in `licenses/OpenVLA-OFT-MIT.txt`.
+
+The source implementation tokenizes instructions with lowercase `[a-z0-9]+` and uses binary presence features. The released weight file preserves every original tensor and is loadable by the public `ResidualAdapter` class. `example.py` applies it to one released matched record on CPU.

README.md

@@ -0,0 +1,76 @@
+---
+language: en
+license: mit
+base_model: openvla/openvla-7b
+base_model_relation: adapter
+tags:
+  - robotics
+  - vision-language-action
+  - residual-adapter
+  - bas-vla
+  - arxiv:2609.23650
+---
+
+# BAS-VLA: OpenVLA residual adapter
+
+[Paper](https://arxiv.org/abs/2609.23650) · [Project](https://nebulis-lab.com/BAS-VLA/) · [Code](https://github.com/NEBULIS-Lab/BAS-VLA)
+
+This release provides a trained text-conditioned residual adapter and matched inputs for the **OpenVLA / LIBERO-10 offline action-calibration setting**, a selected subset of the settings studied in *Beyond Appearance Shifts: Task-Semantic Action Calibration for VLA Models*.
+
+Each record pairs a shared observation with clean, task-preserving, and semantics-changing instructions and their frozen-carrier actions. The included cases cover target-object and executability-condition changes. The adapter applies a bounded action correction using the base action and binary bag-of-words instruction features.
+
+## Files
+
+| Location | Contents |
+| --- | --- |
+| `adapter/` | Trained residual weights, vocabulary, and loading/action metadata |
+| `data/triplets.json` | Matched action triplets, expert actions, and demonstration-level train/validation/test assignments |
+| `data/observations.npz` | Source agent-view RGB images aligned with the records |
+| `data/states/` | Sampled simulator states and source demonstration initial states |
+| `data/scenes/` | The corresponding LIBERO BDDL scenes |
+| `data/cases.json`, `data/manifest.json` | Instruction triplets, source references, array alignment, and input preprocessing |
+| `normalization/` | Action statistics used to construct the frozen OpenVLA actions and their source information |
+| `training_config.json`, `offline_metrics.json` | Training objective configuration and offline evaluation summaries |
+| `example.py` | CPU example applying the adapter to a supplied record |
+
+## Download and use
+
+Install the [BAS-VLA code](https://github.com/NEBULIS-Lab/BAS-VLA) and its dependencies, then download the files:
+
+```python
+from huggingface_hub import snapshot_download
+
+snapshot_download("Shuaijun/BAS-VLA", local_dir="artifacts/bas-vla")
+```
+
+Run the included example from the directory containing `artifacts/`:
+
+```bash
+python artifacts/bas-vla/example.py
+```
+
+The example loads the adapter on CPU and prints calibrated clean, control, and changed-instruction actions for a held-out record. Use `--record-index` to select another record.
+
+## Action and feature conventions
+
+Use the exact contract in `adapter/adapter_metadata.json` when supplying new actions. The carrier is `openvla/openvla-7b`; the cached actions use the bundled `libero_10_no_noops` statistics from `moojink/openvla-7b-oft-finetuned-libero-10`, followed by the documented gripper transform. Instructions use the supplied binary bag-of-words vocabulary.
+
+The checkpoint's action-space identifier is `openvla_libero10_unnormalized_sign_gripper_offline`. The code repository's OpenPI and OpenVLA-OFT rollout runners require adapters matched to their own carrier and action conventions. Load this checkpoint through the supplied example or the `ResidualAdapter` class with the recorded convention.
+
+`data/manifest.json` explains how RGB arrays, simulator states, scene files, and demonstration frames align. LIBERO supplies the scene assets and simulator. Base VLA weights are obtained separately from their upstream model repositories.
+
+## Sources and license
+
+BAS-VLA contributions are released under the MIT License. The included LIBERO-derived data and OpenVLA-OFT normalization statistics retain their upstream MIT notices. See [NOTICE.md](NOTICE.md) and `licenses/` for the source links and notices.
+
+## Citation
+
+```bibtex
+@inproceedings{liu2026basvla,
+  title = {Beyond Appearance Shifts: Task-Semantic Action Calibration for VLA Models},
+  author = {Liu, Shuaijun and You, Feiyang and Wu, Chengyu and Hao, Shuyang and Zhang, Chenglong and Cai, Jingyao and Chen, Xingwei and Su, Ningxin},
+  booktitle = {Advances in Neural Information Processing Systems},
+  year = {2026},
+  url = {https://arxiv.org/abs/2609.23650}
+}
+```

adapter/adapter_metadata.json

@@ -0,0 +1,57 @@
+{
+  "schema_version": 1,
+  "adapter_name": "openvla_libero10_offline_text_action_adapter",
+  "artifact_type": "trained_residual_adapter",
+  "carrier": "openvla/openvla-7b",
+  "input_dim": 28,
+  "hidden_dim": 128,
+  "output_dim": 7,
+  "delta_scale": 0.8,
+  "vocab_size": 21,
+  "feature_type": "binary_bow",
+  "feature_normalization": "binary_presence_no_scaling",
+  "feature_tokenization": "lowercase regex [a-z0-9]+",
+  "feature_source": "explicit vocabulary in vocab.json, built from the two configured instruction triplets",
+  "action_space": "openvla_libero10_unnormalized_sign_gripper_offline",
+  "action_layout": "per_step",
+  "action_normalization": {
+    "carrier_output": "OpenVLA predict_action unnormalization using libero_10_no_noops action statistics",
+    "statistics_file": "../normalization/dataset_statistics.json",
+    "statistics_source": "moojink/openvla-7b-oft-finetuned-libero-10",
+    "continuous_coordinates": "the first six predict_action output coordinates, without additional scaling",
+    "gripper_transform": "g = 2 * g - 1; g = sign(g) if g != 0 else -1",
+    "reference_actions": "original LIBERO HDF5 actions at the recorded demo/frame",
+    "adapter_output": "residual applied to all seven coordinates; no post-adapter clipping or gripper inversion in the offline evaluator"
+  },
+  "objective_config": {
+    "objective_preset": "legacy",
+    "squared_error_reduction": "mean",
+    "lambda_imitation": 1.0,
+    "lambda_expert_margin": 1.0,
+    "hidden_dim": 128,
+    "delta_scale": 0.8,
+    "margin": 0.3,
+    "lambda_consistency": 0.5,
+    "lambda_margin": 2.0,
+    "lambda_anchor": 0.05,
+    "lr": 0.002,
+    "epochs": 100,
+    "batch_size": 64
+  },
+  "seed": null,
+  "seed_provenance": "The original results file omits the run seed; the source trainer parser defaults to 7.",
+  "training_setting": "LIBERO-10 offline action calibration, target-object and executability-condition instruction triplets",
+  "training_timestamp": "2026-04-17 17:27:07",
+  "compatibility": {
+    "implementation": "bas_vla.breaking.residual_adapter.ResidualAdapter",
+    "load_state_dict": "strict; no tensor conversion",
+    "runtime_action_space": "openvla_libero10_unnormalized_sign_gripper_offline",
+    "deployment_contract": "Use only with actions and instruction features following this metadata. The published OpenPI and OpenVLA-OFT runners use different action-space identifiers."
+  },
+  "files": {
+    "weights": "residual_adapter.pt",
+    "vocab": "vocab.json",
+    "records": "../data/triplets.json",
+    "metrics": "../offline_metrics.json"
+  }
+}

adapter/residual_adapter.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cba0b8fd9c53a559145096400664677e443445f1b7e9b7c95fa779e461e03f63
+size 87525

adapter/vocab.json

@@ -0,0 +1,23 @@
+{
+  "put": 0,
+  "both": 1,
+  "the": 2,
+  "alphabet": 3,
+  "soup": 4,
+  "and": 5,
+  "tomato": 6,
+  "sauce": 7,
+  "in": 8,
+  "basket": 9,
+  "butter": 10,
+  "place": 11,
+  "into": 12,
+  "turn": 13,
+  "on": 14,
+  "stove": 15,
+  "moka": 16,
+  "pot": 17,
+  "it": 18,
+  "off": 19,
+  "switch": 20
+}

data/cases.json

@@ -0,0 +1,29 @@
+{
+  "suite": "libero_10",
+  "cases": [
+    {
+      "case_id": "break_target_object",
+      "family": "target_object",
+      "suite": "libero_10",
+      "task_id": 0,
+      "demo_relpath": "libero_10/LIVING_ROOM_SCENE2_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket_demo.hdf5",
+      "clean_instruction": "put both the alphabet soup and the tomato sauce in the basket",
+      "break_instruction": "put both the alphabet soup and the butter in the basket",
+      "control_instruction": "place both the alphabet soup and the tomato sauce into the basket",
+      "changed_semantics": "target object tomato sauce -> butter",
+      "bddl_file": "scenes/LIVING_ROOM_SCENE2_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket.bddl"
+    },
+    {
+      "case_id": "break_executability_condition",
+      "family": "executability_condition",
+      "suite": "libero_10",
+      "task_id": 2,
+      "demo_relpath": "libero_10/KITCHEN_SCENE3_turn_on_the_stove_and_put_the_moka_pot_on_it_demo.hdf5",
+      "clean_instruction": "turn on the stove and put the moka pot on it",
+      "break_instruction": "turn off the stove and put the moka pot on it",
+      "control_instruction": "switch on the stove and put the moka pot on it",
+      "changed_semantics": "turn on -> turn off",
+      "bddl_file": "scenes/KITCHEN_SCENE3_turn_on_the_stove_and_put_the_moka_pot_on_it.bddl"
+    }
+  ]
+}

data/manifest.json

@@ -0,0 +1,120 @@
+{
+  "setting": "OpenVLA / LIBERO-10 offline action calibration",
+  "num_records": 400,
+  "split_counts": {
+    "train": 240,
+    "val": 80,
+    "test": 80
+  },
+  "source_dataset": "https://huggingface.co/datasets/yifengzhu-hf/LIBERO-datasets",
+  "source_code": "https://github.com/Lifelong-Robot-Learning/LIBERO",
+  "records_file": "triplets.json",
+  "observation_file": "observations.npz",
+  "observation_array": "agentview_rgb",
+  "observation_shape": [
+    400,
+    128,
+    128,
+    3
+  ],
+  "observation_dtype": "uint8",
+  "observation_alignment": "triplets.json observation_index is the first-axis index of agentview_rgb; original HDF5 images with no flip, crop or resizing",
+  "paired_inputs": "All three instruction/action pairs in a record share this same source observation and simulator state.",
+  "source_states": {
+    "break_target_object": {
+      "sampled_states_shape": [
+        200,
+        123
+      ],
+      "demo_initial_states_shape": [
+        50,
+        123
+      ],
+      "sampled_states_dtype": "float64"
+    },
+    "break_executability_condition": {
+      "sampled_states_shape": [
+        200,
+        47
+      ],
+      "demo_initial_states_shape": [
+        50,
+        47
+      ],
+      "sampled_states_dtype": "float64"
+    }
+  },
+  "state_alignment": "Per-case NPZ record_indices indexes triplets.json. demo_numbers indexes demo_initial_states. States are source LIBERO simulator snapshots; scene models and simulator conventions come from LIBERO.",
+  "split_policy": {
+    "demo_order": "integer suffix in demo_N",
+    "train_demos_per_case": [
+      0,
+      1,
+      2,
+      3,
+      4,
+      5,
+      6,
+      7,
+      8,
+      9,
+      10,
+      11,
+      12,
+      13,
+      14,
+      15,
+      16,
+      17,
+      18,
+      19,
+      20,
+      21,
+      22,
+      23,
+      24,
+      25,
+      26,
+      27,
+      28,
+      29
+    ],
+    "val_demos_per_case": [
+      30,
+      31,
+      32,
+      33,
+      34,
+      35,
+      36,
+      37,
+      38,
+      39
+    ],
+    "test_demos_per_case": [
+      40,
+      41,
+      42,
+      43,
+      44,
+      45,
+      46,
+      47,
+      48,
+      49
+    ],
+    "frames_per_demo": 4,
+    "frame_rule": "round(linspace(0, number_of_frames - 1, 4))"
+  },
+  "carrier_input_preprocessing": {
+    "source_view": "agentview_rgb",
+    "center_crop_area_fraction": 0.9,
+    "resize": 224,
+    "resample": "PIL LANCZOS",
+    "prompt": "In: What action should the robot take to {instruction.lower()}?\nOut:",
+    "sampling": "do_sample=False",
+    "provenance": "Source generator implementation and model registry; resize=224 is the parser default, and the source results do not store preprocessing CLI overrides."
+  },
+  "expert_reference": "clean LIBERO demonstration action at the same source demo/frame",
+  "license_file": "../licenses/LIBERO-MIT.txt"
+}

data/observations.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7fca8917d6bb9011e0975a3859472901c682f5dc7ed42c3b55be71bcae6f30cd
+size 9108424

data/scenes/KITCHEN_SCENE3_turn_on_the_stove_and_put_the_moka_pot_on_it.bddl

@@ -0,0 +1,68 @@
+(define (problem LIBERO_Kitchen_Tabletop_Manipulation)
+  (:domain robosuite)
+  (:language turn on the stove and put the moka pot on it)
+    (:regions
+      (flat_stove_init_region
+          (:target kitchen_table)
+          (:ranges (
+              (-0.21000000000000002 0.19 -0.19 0.21000000000000002)
+            )
+          )
+          (:yaw_rotation (
+              (0.0 0.0)
+            )
+          )
+      )
+      (frypan_init_region
+          (:target kitchen_table)
+          (:ranges (
+              (-0.07500000000000001 -0.275 -0.025 -0.225)
+            )
+          )
+          (:yaw_rotation (
+              (0.0 0.0)
+            )
+          )
+      )
+      (moka_pot_init_region
+          (:target kitchen_table)
+          (:ranges (
+              (0.025 -0.025 0.07500000000000001 0.025)
+            )
+          )
+          (:yaw_rotation (
+              (0.0 0.0)
+            )
+          )
+      )
+      (cook_region
+          (:target flat_stove_1)
+      )
+    )
+
+  (:fixtures
+    kitchen_table - kitchen_table
+    flat_stove_1 - flat_stove
+  )
+
+  (:objects
+    chefmate_8_frypan_1 - chefmate_8_frypan
+    moka_pot_1 - moka_pot
+  )
+
+  (:obj_of_interest
+    moka_pot_1
+    flat_stove_1
+  )
+
+  (:init
+    (On flat_stove_1 kitchen_table_flat_stove_init_region)
+    (On chefmate_8_frypan_1 kitchen_table_frypan_init_region)
+    (On moka_pot_1 kitchen_table_moka_pot_init_region)
+  )
+
+  (:goal
+    (And (Turnon flat_stove_1) (On moka_pot_1 flat_stove_1_cook_region))
+  )
+
+)

data/scenes/LIVING_ROOM_SCENE2_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket.bddl

@@ -0,0 +1,134 @@
+(define (problem LIBERO_Living_Room_Tabletop_Manipulation)
+  (:domain robosuite)
+  (:language put both the alphabet soup and the tomato sauce in the basket)
+    (:regions
+      (basket_init_region
+          (:target living_room_table)
+          (:ranges (
+              (-0.01 0.25 0.01 0.27)
+            )
+          )
+          (:yaw_rotation (
+              (0.0 0.0)
+            )
+          )
+      )
+      (milk_init_region
+          (:target living_room_table)
+          (:ranges (
+              (0.025 -0.125 0.07500000000000001 -0.07500000000000001)
+            )
+          )
+          (:yaw_rotation (
+              (0.0 0.0)
+            )
+          )
+      )
+      (cream_cheese_init_region
+          (:target living_room_table)
+          (:ranges (
+              (0.07500000000000001 -0.225 0.125 -0.17500000000000002)
+            )
+          )
+          (:yaw_rotation (
+              (0.0 0.0)
+            )
+          )
+      )
+      (orange_juice_init_region
+          (:target living_room_table)
+          (:ranges (
+              (-0.025 -0.275 0.025 -0.225)
+            )
+          )
+          (:yaw_rotation (
+              (0.0 0.0)
+            )
+          )
+      )
+      (tomato_sauce_init_region
+          (:target living_room_table)
+          (:ranges (
+              (-0.125 0.025 -0.07500000000000001 0.07500000000000001)
+            )
+          )
+          (:yaw_rotation (
+              (0.0 0.0)
+            )
+          )
+      )
+      (alphabet_soup_init_region
+          (:target living_room_table)
+          (:ranges (
+              (-0.125 -0.175 -0.07500000000000001 -0.125)
+            )
+          )
+          (:yaw_rotation (
+              (0.0 0.0)
+            )
+          )
+      )
+      (butter_init_region
+          (:target living_room_table)
+          (:ranges (
+              (0.025 0.025 0.07500000000000001 0.07500000000000001)
+            )
+          )
+          (:yaw_rotation (
+              (0.0 0.0)
+            )
+          )
+      )
+      (ketchup_init_region
+          (:target living_room_table)
+          (:ranges (
+              (-0.275 -0.175 -0.225 -0.125)
+            )
+          )
+          (:yaw_rotation (
+              (0.0 0.0)
+            )
+          )
+      )
+      (contain_region
+          (:target basket_1)
+      )
+    )
+
+  (:fixtures
+    living_room_table - living_room_table
+  )
+
+  (:objects
+    alphabet_soup_1 - alphabet_soup
+    cream_cheese_1 - cream_cheese
+    tomato_sauce_1 - tomato_sauce
+    ketchup_1 - ketchup
+    orange_juice_1 - orange_juice
+    milk_1 - milk
+    butter_1 - butter
+    basket_1 - basket
+  )
+
+  (:obj_of_interest
+    alphabet_soup_1
+    tomato_sauce_1
+    basket_1
+  )
+
+  (:init
+    (On alphabet_soup_1 living_room_table_alphabet_soup_init_region)
+    (On cream_cheese_1 living_room_table_cream_cheese_init_region)
+    (On tomato_sauce_1 living_room_table_tomato_sauce_init_region)
+    (On ketchup_1 living_room_table_ketchup_init_region)
+    (On milk_1 living_room_table_milk_init_region)
+    (On orange_juice_1 living_room_table_orange_juice_init_region)
+    (On butter_1 living_room_table_butter_init_region)
+    (On basket_1 living_room_table_basket_init_region)
+  )
+
+  (:goal
+    (And (In alphabet_soup_1 basket_1_contain_region) (In tomato_sauce_1 basket_1_contain_region))
+  )
+
+)

data/states/break_executability_condition.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fa5b77875f2a4292fb198c8fc4c65c9cb2d6638ed4090d2f0e7e6add8964595
+size 72539

data/states/break_target_object.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ceb28389d3c8066c542b68dba96d1002d58bbc9a1179f23cd6fd4f69c3582758
+size 170568

example.py

@@ -0,0 +1,55 @@
+"""Apply the released adapter to one matched offline record on CPU.
+
+Install the BAS-VLA code repository and its dependencies first. The input action
+contract is recorded in adapter/adapter_metadata.json. Use that exact contract
+when supplying new actions from a carrier.
+"""
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+import numpy as np
+import torch
+
+from bas_vla.breaking.residual_adapter import ResidualAdapter, encode_instruction_bow
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--artifact-dir", type=Path, default=Path(__file__).resolve().parent)
+    parser.add_argument("--record-index", type=int, default=160, help="Default: first held-out test record.")
+    args = parser.parse_args()
+    root = args.artifact_dir
+    meta = json.loads((root / "adapter/adapter_metadata.json").read_text())
+    vocab = json.loads((root / "adapter/vocab.json").read_text())
+    records = json.loads((root / "data/triplets.json").read_text())
+    if not 0 <= args.record_index < len(records):
+        parser.error("record-index is outside the released records")
+    record = records[args.record_index]
+    model = ResidualAdapter(
+        input_dim=meta["input_dim"], hidden_dim=meta["hidden_dim"],
+        output_dim=meta["output_dim"], delta_scale=meta["delta_scale"],
+    )
+    weights = torch.load(root / "adapter/residual_adapter.pt", map_location="cpu", weights_only=True)
+    model.load_state_dict(weights, strict=True)
+    model.eval()
+    outputs = {}
+    with torch.no_grad():
+        for variant in ("clean", "control", "break"):
+            action = np.asarray(record[f"{variant}_base_action"], dtype=np.float32)
+            features = encode_instruction_bow(record[f"{variant}_instruction"], vocab)
+            calibrated = model(torch.from_numpy(action[None]), torch.from_numpy(features[None]))[0]
+            outputs[variant] = {
+                "instruction": record[f"{variant}_instruction"],
+                "base_action": action.tolist(), "calibrated_action": calibrated.tolist(),
+            }
+    print(json.dumps({
+        "record_id": record["record_id"], "split": record["split"],
+        "action_space": meta["action_space"], "outputs": outputs,
+    }, indent=2))
+
+
+if __name__ == "__main__":
+    main()

licenses/LIBERO-MIT.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 Lifelong Robot Learning
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

licenses/OpenVLA-OFT-MIT.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 Moo Jin Kim, Chelsea Finn, Percy Liang.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

normalization/dataset_statistics.json

@@ -0,0 +1,133 @@
+{
+  "libero_10_no_noops": {
+    "action": {
+      "mean": [
+        0.01820324920117855,
+        0.05858374014496803,
+        -0.05592384561896324,
+        0.004626928828656673,
+        0.00289608770981431,
+        -0.007673131301999092,
+        0.5457824468612671
+      ],
+      "std": [
+        0.2825464606285095,
+        0.35904666781425476,
+        0.3673802614212036,
+        0.03770702704787254,
+        0.05429719388484955,
+        0.08725254982709885,
+        0.49815231561660767
+      ],
+      "max": [
+        0.9375,
+        0.9375,
+        0.9375,
+        0.30000001192092896,
+        0.29357144236564636,
+        0.375,
+        1.0
+      ],
+      "min": [
+        -0.9375,
+        -0.9375,
+        -0.9375,
+        -0.23642857372760773,
+        -0.3053571283817291,
+        -0.3675000071525574,
+        0.0
+      ],
+      "q01": [
+        -0.6348214149475098,
+        -0.7741071581840515,
+        -0.7633928656578064,
+        -0.09749999642372131,
+        -0.14819999992847435,
+        -0.2742857038974762,
+        0.0
+      ],
+      "q99": [
+        0.7714285850524902,
+        0.8464285731315613,
+        0.9375,
+        0.13928571343421936,
+        0.15964286029338837,
+        0.3246428668498993,
+        1.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        false
+      ]
+    },
+    "proprio": {
+      "mean": [
+        -0.04190658777952194,
+        0.03539430722594261,
+        0.8257141709327698,
+        2.908308267593384,
+        -0.5562185049057007,
+        -0.16649018228054047,
+        0.028316624462604523,
+        -0.028561657294631004
+      ],
+      "std": [
+        0.10743364691734314,
+        0.14424669742584229,
+        0.2572328448295593,
+        0.3441362977027893,
+        1.234421730041504,
+        0.3579835891723633,
+        0.013308707624673843,
+        0.013174631632864475
+      ],
+      "max": [
+        0.21031762659549713,
+        0.39128610491752625,
+        1.3332009315490723,
+        3.6714255809783936,
+        3.560650587081909,
+        1.386339545249939,
+        0.04160946607589722,
+        0.0013633022317662835
+      ],
+      "min": [
+        -0.4828203022480011,
+        -0.3255046010017395,
+        0.445506751537323,
+        1.1321442127227783,
+        -3.641430377960205,
+        -1.842738389968872,
+        -0.0010040868073701859,
+        -0.04111652821302414
+      ],
+      "q01": [
+        -0.3899900782108307,
+        -0.2838300323486328,
+        0.44795057058334353,
+        1.8810229921340942,
+        -2.886677579879761,
+        -1.1599004411697387,
+        0.002066459748893976,
+        -0.04001387819647789
+      ],
+      "q99": [
+        0.1530261474847791,
+        0.32915401458740223,
+        1.2546923208236693,
+        3.303542451858519,
+        2.7496529006957933,
+        0.6893712210655194,
+        0.040048558115959164,
+        -0.0017598449345678235
+      ]
+    },
+    "num_transitions": 101469,
+    "num_trajectories": 379
+  }
+}

normalization/source.json

@@ -0,0 +1,9 @@
+{
+  "model": "moojink/openvla-7b-oft-finetuned-libero-10",
+  "file": "dataset_statistics.json",
+  "key": "libero_10_no_noops",
+  "url": "https://huggingface.co/moojink/openvla-7b-oft-finetuned-libero-10/blob/main/dataset_statistics.json",
+  "purpose": "Exact external statistics used by the source OpenVLA-base record generator",
+  "license": "MIT",
+  "license_file": "../licenses/OpenVLA-OFT-MIT.txt"
+}

offline_metrics.json

@@ -0,0 +1,370 @@
+{
+  "evaluation_type": "offline paired-action metrics",
+  "timestamp": "2026-04-17 17:27:07",
+  "num_records": {
+    "train": 240,
+    "val": 80,
+    "test": 80
+  },
+  "models": {
+    "identity": {
+      "train_metrics": {
+        "mean_break_clean_l2": 0.7432431578636169,
+        "mean_break_expert_l2": 1.4091085195541382,
+        "mean_clean_mse": 0.3958700895309448,
+        "mean_control_clean_l2": 0.7472800016403198,
+        "mean_control_expert_l2": 1.4053751230239868,
+        "mean_control_mse": 0.37037795782089233,
+        "break_gt_control_rate_clean_ref": 0.5458333333333333,
+        "break_gt_control_rate_expert_ref": 0.4875,
+        "mean_clean_ref_gap": -0.004036815837025642,
+        "mean_expert_ref_gap": 0.003733340883627534,
+        "by_case": {
+          "break_target_object": {
+            "mean_break_clean_l2": 0.543735682964325,
+            "mean_break_expert_l2": 1.5344077348709106,
+            "mean_clean_mse": 0.45629745721817017,
+            "mean_control_clean_l2": 0.6087575554847717,
+            "mean_control_expert_l2": 1.5666840076446533,
+            "mean_control_mse": 0.42383497953414917,
+            "break_gt_control_rate_clean_ref": 0.4666666666666667,
+            "break_gt_control_rate_expert_ref": 0.49166666666666664,
+            "mean_clean_ref_gap": -0.06502188742160797,
+            "mean_expert_ref_gap": -0.032276351004838943
+          },
+          "break_executability_condition": {
+            "mean_break_clean_l2": 0.9427506923675537,
+            "mean_break_expert_l2": 1.2838094234466553,
+            "mean_clean_mse": 0.3354426920413971,
+            "mean_control_clean_l2": 0.8858024477958679,
+            "mean_control_expert_l2": 1.2440663576126099,
+            "mean_control_mse": 0.3169209063053131,
+            "break_gt_control_rate_clean_ref": 0.625,
+            "break_gt_control_rate_expert_ref": 0.48333333333333334,
+            "mean_clean_ref_gap": 0.056948255747556686,
+            "mean_expert_ref_gap": 0.039743032306432724
+          }
+        },
+        "num_records": 240
+      },
+      "val_metrics": {
+        "mean_break_clean_l2": 0.6865259408950806,
+        "mean_break_expert_l2": 1.3638770580291748,
+        "mean_clean_mse": 0.35426950454711914,
+        "mean_control_clean_l2": 0.6174260377883911,
+        "mean_control_expert_l2": 1.4026581048965454,
+        "mean_control_mse": 0.3633658289909363,
+        "break_gt_control_rate_clean_ref": 0.575,
+        "break_gt_control_rate_expert_ref": 0.4375,
+        "mean_clean_ref_gap": 0.06909994781017303,
+        "mean_expert_ref_gap": -0.03878090903162956,
+        "by_case": {
+          "break_target_object": {
+            "mean_break_clean_l2": 0.6781854629516602,
+            "mean_break_expert_l2": 1.4673945903778076,
+            "mean_clean_mse": 0.43167275190353394,
+            "mean_control_clean_l2": 0.5327732563018799,
+            "mean_control_expert_l2": 1.5085182189941406,
+            "mean_control_mse": 0.3970142602920532,
+            "break_gt_control_rate_clean_ref": 0.575,
+            "break_gt_control_rate_expert_ref": 0.375,
+            "mean_clean_ref_gap": 0.14541222155094147,
+            "mean_expert_ref_gap": -0.041123662143945694
+          },
+          "break_executability_condition": {
+            "mean_break_clean_l2": 0.694866418838501,
+            "mean_break_expert_l2": 1.2603596448898315,
+            "mean_clean_mse": 0.27686622738838196,
+            "mean_control_clean_l2": 0.7020787596702576,
+            "mean_control_expert_l2": 1.2967978715896606,
+            "mean_control_mse": 0.32971736788749695,
+            "break_gt_control_rate_clean_ref": 0.575,
+            "break_gt_control_rate_expert_ref": 0.5,
+            "mean_clean_ref_gap": -0.007212340831756592,
+            "mean_expert_ref_gap": -0.036438144743442535
+          }
+        },
+        "num_records": 80
+      },
+      "test_metrics": {
+        "mean_break_clean_l2": 0.6800279021263123,
+        "mean_break_expert_l2": 1.6663868427276611,
+        "mean_clean_mse": 0.39530259370803833,
+        "mean_control_clean_l2": 0.7465093731880188,
+        "mean_control_expert_l2": 1.351813554763794,
+        "mean_control_mse": 0.3437272906303406,
+        "break_gt_control_rate_clean_ref": 0.5125,
+        "break_gt_control_rate_expert_ref": 0.5625,
+        "mean_clean_ref_gap": -0.06648143380880356,
+        "mean_expert_ref_gap": 0.31457358598709106,
+        "by_case": {
+          "break_target_object": {
+            "mean_break_clean_l2": 0.4548130929470062,
+            "mean_break_expert_l2": 1.7907543182373047,
+            "mean_clean_mse": 0.4607141613960266,
+            "mean_control_clean_l2": 0.6214543581008911,
+            "mean_control_expert_l2": 1.5308020114898682,
+            "mean_control_mse": 0.40925103425979614,
+            "break_gt_control_rate_clean_ref": 0.4,
+            "break_gt_control_rate_expert_ref": 0.55,
+            "mean_clean_ref_gap": -0.1666412204504013,
+            "mean_expert_ref_gap": 0.25995248556137085
+          },
+          "break_executability_condition": {
+            "mean_break_clean_l2": 0.9052427411079407,
+            "mean_break_expert_l2": 1.5420197248458862,
+            "mean_clean_mse": 0.32989105582237244,
+            "mean_control_clean_l2": 0.8715643882751465,
+            "mean_control_expert_l2": 1.1728250980377197,
+            "mean_control_mse": 0.27820348739624023,
+            "break_gt_control_rate_clean_ref": 0.625,
+            "break_gt_control_rate_expert_ref": 0.575,
+            "mean_clean_ref_gap": 0.033678364008665085,
+            "mean_expert_ref_gap": 0.3691946864128113
+          }
+        },
+        "num_records": 80
+      }
+    },
+    "action_only_adapter": {
+      "best_val_loss_total": 1.5956528186798096,
+      "train_metrics": {
+        "mean_break_clean_l2": 0.36038321256637573,
+        "mean_break_expert_l2": 1.1645487546920776,
+        "mean_clean_mse": 0.20006108283996582,
+        "mean_control_clean_l2": 0.2829788327217102,
+        "mean_control_expert_l2": 1.1402603387832642,
+        "mean_control_mse": 0.19075056910514832,
+        "break_gt_control_rate_clean_ref": 0.5875,
+        "break_gt_control_rate_expert_ref": 0.5291666666666667,
+        "mean_clean_ref_gap": 0.07740439474582672,
+        "mean_expert_ref_gap": 0.024288399145007133,
+        "by_case": {
+          "break_target_object": {
+            "mean_break_clean_l2": 0.25087252259254456,
+            "mean_break_expert_l2": 1.198777675628662,
+            "mean_clean_mse": 0.21315599977970123,
+            "mean_control_clean_l2": 0.22465568780899048,
+            "mean_control_expert_l2": 1.1870900392532349,
+            "mean_control_mse": 0.20506463944911957,
+            "break_gt_control_rate_clean_ref": 0.5083333333333333,
+            "break_gt_control_rate_expert_ref": 0.55,
+            "mean_clean_ref_gap": 0.026216844096779823,
+            "mean_expert_ref_gap": 0.011687582358717918
+          },
+          "break_executability_condition": {
+            "mean_break_clean_l2": 0.4698939025402069,
+            "mean_break_expert_l2": 1.1303197145462036,
+            "mean_clean_mse": 0.1869661659002304,
+            "mean_control_clean_l2": 0.34130197763442993,
+            "mean_control_expert_l2": 1.0934306383132935,
+            "mean_control_mse": 0.17643649876117706,
+            "break_gt_control_rate_clean_ref": 0.6666666666666666,
+            "break_gt_control_rate_expert_ref": 0.5083333333333333,
+            "mean_clean_ref_gap": 0.12859195470809937,
+            "mean_expert_ref_gap": 0.0368892140686512
+          }
+        },
+        "num_records": 240
+      },
+      "val_metrics": {
+        "mean_break_clean_l2": 0.34965047240257263,
+        "mean_break_expert_l2": 1.1548038721084595,
+        "mean_clean_mse": 0.1989673674106598,
+        "mean_control_clean_l2": 0.2915192246437073,
+        "mean_control_expert_l2": 1.1498476266860962,
+        "mean_control_mse": 0.19426663219928741,
+        "break_gt_control_rate_clean_ref": 0.6125,
+        "break_gt_control_rate_expert_ref": 0.55,
+        "mean_clean_ref_gap": 0.05813126638531685,
+        "mean_expert_ref_gap": 0.004956287331879139,
+        "by_case": {
+          "break_target_object": {
+            "mean_break_clean_l2": 0.2960914373397827,
+            "mean_break_expert_l2": 1.1769672632217407,
+            "mean_clean_mse": 0.2139633595943451,
+            "mean_control_clean_l2": 0.262840211391449,
+            "mean_control_expert_l2": 1.1790968179702759,
+            "mean_control_mse": 0.2026643306016922,
+            "break_gt_control_rate_clean_ref": 0.575,
+            "break_gt_control_rate_expert_ref": 0.5,
+            "mean_clean_ref_gap": 0.03325119987130165,
+            "mean_expert_ref_gap": -0.002129593398422003
+          },
+          "break_executability_condition": {
+            "mean_break_clean_l2": 0.4032095968723297,
+            "mean_break_expert_l2": 1.1326402425765991,
+            "mean_clean_mse": 0.1839713752269745,
+            "mean_control_clean_l2": 0.3201982378959656,
+            "mean_control_expert_l2": 1.1205980777740479,
+            "mean_control_mse": 0.18586888909339905,
+            "break_gt_control_rate_clean_ref": 0.65,
+            "break_gt_control_rate_expert_ref": 0.6,
+            "mean_clean_ref_gap": 0.08301133662462234,
+            "mean_expert_ref_gap": 0.012042166665196419
+          }
+        },
+        "num_records": 80
+      },
+      "test_metrics": {
+        "mean_break_clean_l2": 0.3397909700870514,
+        "mean_break_expert_l2": 1.2354549169540405,
+        "mean_clean_mse": 0.21382936835289001,
+        "mean_control_clean_l2": 0.3245377540588379,
+        "mean_control_expert_l2": 1.1614665985107422,
+        "mean_control_mse": 0.19767430424690247,
+        "break_gt_control_rate_clean_ref": 0.5375,
+        "break_gt_control_rate_expert_ref": 0.5125,
+        "mean_clean_ref_gap": 0.015253203921020031,
+        "mean_expert_ref_gap": 0.07398830354213715,
+        "by_case": {
+          "break_target_object": {
+            "mean_break_clean_l2": 0.23053522408008575,
+            "mean_break_expert_l2": 1.2631430625915527,
+            "mean_clean_mse": 0.22174766659736633,
+            "mean_control_clean_l2": 0.25600486993789673,
+            "mean_control_expert_l2": 1.2138757705688477,
+            "mean_control_mse": 0.21398238837718964,
+            "break_gt_control_rate_clean_ref": 0.45,
+            "break_gt_control_rate_expert_ref": 0.475,
+            "mean_clean_ref_gap": -0.025469651445746422,
+            "mean_expert_ref_gap": 0.04926721006631851
+          },
+          "break_executability_condition": {
+            "mean_break_clean_l2": 0.44904670119285583,
+            "mean_break_expert_l2": 1.2077667713165283,
+            "mean_clean_mse": 0.2059110701084137,
+            "mean_control_clean_l2": 0.39307063817977905,
+            "mean_control_expert_l2": 1.1090574264526367,
+            "mean_control_mse": 0.1813661754131317,
+            "break_gt_control_rate_clean_ref": 0.625,
+            "break_gt_control_rate_expert_ref": 0.55,
+            "mean_clean_ref_gap": 0.05597605183720589,
+            "mean_expert_ref_gap": 0.09870939701795578
+          }
+        },
+        "num_records": 80
+      }
+    },
+    "text_action_adapter": {
+      "best_val_loss_total": 0.36617913842201233,
+      "train_metrics": {
+        "mean_break_clean_l2": 1.9662588834762573,
+        "mean_break_expert_l2": 2.3562726974487305,
+        "mean_clean_mse": 0.1834717094898224,
+        "mean_control_clean_l2": 0.2672695815563202,
+        "mean_control_expert_l2": 1.0719788074493408,
+        "mean_control_mse": 0.1756630837917328,
+        "break_gt_control_rate_clean_ref": 1.0,
+        "break_gt_control_rate_expert_ref": 1.0,
+        "mean_clean_ref_gap": 1.6989891529083252,
+        "mean_expert_ref_gap": 1.2842938899993896,
+        "by_case": {
+          "break_target_object": {
+            "mean_break_clean_l2": 1.7676277160644531,
+            "mean_break_expert_l2": 2.259204864501953,
+            "mean_clean_mse": 0.2054469883441925,
+            "mean_control_clean_l2": 0.2036994993686676,
+            "mean_control_expert_l2": 1.1436399221420288,
+            "mean_control_mse": 0.19497980177402496,
+            "break_gt_control_rate_clean_ref": 1.0,
+            "break_gt_control_rate_expert_ref": 1.0,
+            "mean_clean_ref_gap": 1.5639280080795288,
+            "mean_expert_ref_gap": 1.1155647039413452
+          },
+          "break_executability_condition": {
+            "mean_break_clean_l2": 2.1648900508880615,
+            "mean_break_expert_l2": 2.4533402919769287,
+            "mean_clean_mse": 0.16149643063545227,
+            "mean_control_clean_l2": 0.3308396339416504,
+            "mean_control_expert_l2": 1.0003175735473633,
+            "mean_control_mse": 0.1563463807106018,
+            "break_gt_control_rate_clean_ref": 1.0,
+            "break_gt_control_rate_expert_ref": 1.0,
+            "mean_clean_ref_gap": 1.8340502977371216,
+            "mean_expert_ref_gap": 1.453023076057434
+          }
+        },
+        "num_records": 240
+      },
+      "val_metrics": {
+        "mean_break_clean_l2": 1.9402227401733398,
+        "mean_break_expert_l2": 2.3244988918304443,
+        "mean_clean_mse": 0.168838232755661,
+        "mean_control_clean_l2": 0.23467329144477844,
+        "mean_control_expert_l2": 1.0611259937286377,
+        "mean_control_mse": 0.17241854965686798,
+        "break_gt_control_rate_clean_ref": 1.0,
+        "break_gt_control_rate_expert_ref": 1.0,
+        "mean_clean_ref_gap": 1.7055492401123047,
+        "mean_expert_ref_gap": 1.2633731365203857,
+        "by_case": {
+          "break_target_object": {
+            "mean_break_clean_l2": 1.78369140625,
+            "mean_break_expert_l2": 2.2142577171325684,
+            "mean_clean_mse": 0.19796983897686005,
+            "mean_control_clean_l2": 0.19418850541114807,
+            "mean_control_expert_l2": 1.1197407245635986,
+            "mean_control_mse": 0.18957313895225525,
+            "break_gt_control_rate_clean_ref": 1.0,
+            "break_gt_control_rate_expert_ref": 1.0,
+            "mean_clean_ref_gap": 1.5895028114318848,
+            "mean_expert_ref_gap": 1.0945169925689697
+          },
+          "break_executability_condition": {
+            "mean_break_clean_l2": 2.0967538356781006,
+            "mean_break_expert_l2": 2.4347403049468994,
+            "mean_clean_mse": 0.13970664143562317,
+            "mean_control_clean_l2": 0.2751580476760864,
+            "mean_control_expert_l2": 1.0025112628936768,
+            "mean_control_mse": 0.15526394546031952,
+            "break_gt_control_rate_clean_ref": 1.0,
+            "break_gt_control_rate_expert_ref": 1.0,
+            "mean_clean_ref_gap": 1.8215957880020142,
+            "mean_expert_ref_gap": 1.4322290420532227
+          }
+        },
+        "num_records": 80
+      },
+      "test_metrics": {
+        "mean_break_clean_l2": 1.9752228260040283,
+        "mean_break_expert_l2": 2.467482089996338,
+        "mean_clean_mse": 0.18618619441986084,
+        "mean_control_clean_l2": 0.273476779460907,
+        "mean_control_expert_l2": 1.0540869235992432,
+        "mean_control_mse": 0.1706114113330841,
+        "break_gt_control_rate_clean_ref": 1.0,
+        "break_gt_control_rate_expert_ref": 1.0,
+        "mean_clean_ref_gap": 1.7017459869384766,
+        "mean_expert_ref_gap": 1.4133949279785156,
+        "by_case": {
+          "break_target_object": {
+            "mean_break_clean_l2": 1.7627042531967163,
+            "mean_break_expert_l2": 2.334259033203125,
+            "mean_clean_mse": 0.21386170387268066,
+            "mean_control_clean_l2": 0.2173435240983963,
+            "mean_control_expert_l2": 1.1480493545532227,
+            "mean_control_mse": 0.19736817479133606,
+            "break_gt_control_rate_clean_ref": 1.0,
+            "break_gt_control_rate_expert_ref": 1.0,
+            "mean_clean_ref_gap": 1.5453606843948364,
+            "mean_expert_ref_gap": 1.1862096786499023
+          },
+          "break_executability_condition": {
+            "mean_break_clean_l2": 2.18774151802063,
+            "mean_break_expert_l2": 2.6007046699523926,
+            "mean_clean_mse": 0.1585106998682022,
+            "mean_control_clean_l2": 0.3296099901199341,
+            "mean_control_expert_l2": 0.9601246118545532,
+            "mean_control_mse": 0.14385461807250977,
+            "break_gt_control_rate_clean_ref": 1.0,
+            "break_gt_control_rate_expert_ref": 1.0,
+            "mean_clean_ref_gap": 1.8581314086914062,
+            "mean_expert_ref_gap": 1.640580415725708
+          }
+        },
+        "num_records": 80
+      }
+    }
+  }
+}

training_config.json

@@ -0,0 +1,17 @@
+{
+  "objective_preset": "legacy",
+  "overrides": {
+    "squared_error_reduction": "mean",
+    "lambda_imitation": 1.0,
+    "lambda_expert_margin": 1.0,
+    "hidden_dim": 128,
+    "delta_scale": 0.8,
+    "margin": 0.3,
+    "lambda_consistency": 0.5,
+    "lambda_margin": 2.0,
+    "lambda_anchor": 0.05,
+    "lr": 0.002,
+    "epochs": 100,
+    "batch_size": 64
+  }
+}