Restore the original separate English and Chinese dataset cards

README.md

@@ -19,44 +19,43 @@ tags:
 
 **English** · [简体中文](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/README_zh.md)
 
-Expert demonstrations for **robotic value models**, covering **real-world** and **simulation** manipulation. Each domain offers two demonstration settings: **Full-Shot** for training or fine-tuning, and **One-Shot** for reference conditioning or task-specific adaptation.
+Expert demonstrations for **robotic value models**, covering real-world and simulated manipulation tasks. Each domain has two demonstration-access settings: **Full-Shot** for training or fine-tuning, and **One-Shot** for a single task-specific reference or adaptation example.
 
-**LeRobot Dataset v3.0** is the shared release format. Full-Shot means **100 episodes per task**; One-Shot means **one episode per task**, selected from the corresponding Full-Shot pool—not an additional collection.
+The common release format is **LeRobot Dataset v3.0**: per-step Parquet data, camera videos, and task/episode metadata. Robot platforms, cameras, and state/action schemas are specified separately for each subset.
 
-[Preview](#1-demonstration-preview) · [Real-world](#2-real-world) · [Simulation](#3-simulation) · [Quick start](#4-quick-start) · [Format](#5-format-and-repository) · [Selection protocol](#6-collection-and-selection)
+**Available now: Real-World Full-Shot — 2,000 episodes across 20 tasks.** The other three subsets have reserved directories and defined sizes, but do not yet contain trajectory files.
 
-## 1. Demonstration Preview
+[Real-world data](#2-real-world) · [Simulation](#3-simulation) · [Quick start](#5-download-and-reading) · [Repository files](#4-repository-layout)
 
-<video controls muted playsinline preload="none" width="640" height="400" poster="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/resolve/main/assets/previews/real-world-task-00.jpg">
-  <source src="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/resolve/main/assets/previews/real-world-task-00.mp4" type="video/mp4">
-</video>
+## 1. Dataset Overview
 
-**Transfer Blocks Using Plate** · task 0 · episode 0 · head camera · 23.17 seconds. [Open the video](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/resolve/main/assets/previews/real-world-task-00.mp4).
+| Domain | Tasks | Full-Shot | One-Shot |
+| --- | ---: | ---: | ---: |
+| [Real-World](#2-real-world) | 20 | **2,000** | 20<sup>†</sup> |
+| [Simulation](#3-simulation) | 15 | 1,500<sup>†</sup> | 15<sup>†</sup> |
 
-This is a complete episode from the released real-world training data. A simulation preview will accompany that domain's release; no synthetic or unrelated demo is shown here.
+Counts are **episodes**, not video files or individual camera views. Full-Shot provides **100 episodes per task**; One-Shot selects **one episode per task from the same Full-Shot pool**, rather than adding independently collected demonstrations. † denotes the defined size of a reserved subset.
+
+**Which setting should I use?**
+
+- **Full-Shot:** learn or fine-tune task-conditioned value/progress representations using the full demonstration collection.
+- **One-Shot:** condition a model on one reference demonstration, or perform task-specific adaptation under a one-example budget.
+
+This repository contains **standard in-domain training demonstrations**, not the paper's diagnostic or OOD evaluation trajectories. It does **not** supply extra scalar reward or value-target labels; your training method must define its own supervision or target construction.
 
 ## 2. Real-World
 
-<img src="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/resolve/main/assets/domain-real-world.svg" width="24" height="24" alt=""> **G2A · physical robot demonstrations**
+### 2.1 Overview
 
-**20 manipulation tasks**, covering object placement and sorting, household organization, drawers, repeated presses, kitchen activities, and packaging. Choose this domain for value models grounded in real visual observations and physical interaction.
+Real robot demonstrations collected with **G2A**, covering **20 manipulation tasks**. The tasks include object placement and sorting, household organization, drawers, repeated presses, kitchen activities, and packaging. Choose this domain when you need value-model training or reference demonstrations grounded in real visual observations and physical interaction.
 
-<table width="100%">
-  <thead>
-    <tr><th scope="col" width="20%">Setting</th><th scope="col" width="40%">Full-Shot</th><th scope="col" width="40%">One-Shot</th></tr>
-  </thead>
-  <tbody>
-    <tr><th scope="row">Tasks</th><td>20</td><td>20</td></tr>
-    <tr><th scope="row">Episodes</th><td><strong>2,000</strong><br>100 per task</td><td><strong>20</strong><br>1 per task</td></tr>
-    <tr><th scope="row">Download size</th><td><strong>≈23.23 GB</strong></td><td>To be specified</td></tr>
-    <tr><th scope="row">Use for</th><td>Training / fine-tuning</td><td>Reference / adaptation</td></tr>
-    <tr><th scope="row">Availability</th><td><strong>Available</strong></td><td>Reserved</td></tr>
-    <tr><th scope="row">Open</th><td><a href="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/full-shot">Browse files</a> · <a href="#4-quick-start">Quick start</a></td><td><a href="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/one-shot">Subset directory</a> · <a href="#6-collection-and-selection">Selection protocol</a></td></tr>
-  </tbody>
-</table>
+| Setting | Per task | Episodes | Data |
+| --- | ---: | ---: | --- |
+| [**Full-Shot**](#22-full-shot) | 100 | **2,000** | [Files](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/full-shot) |
+| [One-Shot](#23-one-shot) | 1 | 20 | [Reserved](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/one-shot) |
 
 <details>
-<summary><strong>Real-world task catalog: all 20 tasks and short descriptions</strong></summary>
+<summary><strong>Task catalog: all 20 tasks and short descriptions</strong></summary>
 
 Indices match the training data's `task_index`. Display names and descriptions follow Appendix D of the RoboValue paper; the original task strings in `meta/tasks.parquet` are retained.
 
@@ -85,81 +84,112 @@ Indices match the training data's `task_index`. Display names and descriptions f
 
 </details>
 
-## 3. Simulation
+### 2.2 Full-Shot
 
-<img src="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/resolve/main/assets/domain-simulation.svg" width="24" height="24" alt=""> **Simulated manipulation · 15 tasks**
+**Use for training or fine-tuning with all 100 demonstrations per task.** This is the currently available real-world training subset.
 
-For value-model experiments and reference conditioning in simulated environments. Both settings follow the same LeRobot v3.0 release specification; the robot, cameras, and state/action schema will be documented for this domain rather than inferred from G2A.
+[Download and read](#5-download-and-reading) · [Browse files](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/full-shot) · [Feature definitions](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/full-shot/meta/info.json)
 
-<table width="100%">
-  <thead>
-    <tr><th scope="col" width="20%">Setting</th><th scope="col" width="40%">Full-Shot</th><th scope="col" width="40%">One-Shot</th></tr>
-  </thead>
-  <tbody>
-    <tr><th scope="row">Tasks</th><td>15</td><td>15</td></tr>
-    <tr><th scope="row">Episodes</th><td><strong>1,500</strong><br>100 per task</td><td><strong>15</strong><br>1 per task</td></tr>
-    <tr><th scope="row">Download size</th><td>To be specified</td><td>To be specified</td></tr>
-    <tr><th scope="row">Use for</th><td>Training / fine-tuning</td><td>Reference / adaptation</td></tr>
-    <tr><th scope="row">Availability</th><td>Reserved</td><td>Reserved</td></tr>
-    <tr><th scope="row">Open</th><td><a href="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/full-shot">Browse files</a></td><td><a href="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/one-shot">Subset directory</a> · <a href="#6-collection-and-selection">Selection protocol</a></td></tr>
-  </tbody>
-</table>
+**Key statistics**
 
-**Reading the counts:** only real-world Full-Shot currently contains trajectory files. Reserved counts describe the defined scope, not downloadable data. One-Shot reuses Full-Shot episodes, so do not add its counts when reporting unique demonstrations; unreleased download sizes are not estimated.
+| Item | Value |
+| --- | --- |
+| Coverage | **20 tasks × 100 episodes = 2,000 episodes** |
+| Time steps / sampling | **1,530,987** at **30 FPS** |
+| Trajectory duration | **≈14.18 hours** total; **≈25.52 seconds** per episode on average |
+| Cameras | **3 RGB views:** head, left wrist, right wrist |
+| Resolution | Head **640 × 400**; each wrist **640 × 528** |
+| Storage | **≈23.23 GB**; Parquet + HEVC video |
+| Split | `train: 0:2000` |
 
+Duration and time steps count the trajectory once, not three times for the cameras.
 
-## 4. Quick Start
+**What does one trajectory contain?**
 
-**Load the available real-world Full-Shot subset.** Public downloads need no token; the complete subset, including videos, is approximately **23.23 GB**.
+An **episode** is a complete task demonstration: task/episode identifiers, per-step robot observations and actions, and three camera streams. Episode metadata records its length, data-row boundaries, and video segments. In v3.0, episodes may share chunk files; a file is not necessarily one episode.
 
-Python **3.10+**; reader verified with **LeRobot 0.4.2** and **PyAV 15.1.0**. **Dataset v3.0 is the format version, not a software package named LeRobot 3.0.**
+| Signal | Representation |
+| --- | --- |
+| `action` | **52 float32 values:** joint targets **and** end-effector target poses, plus gripper, head, waist, and base-control fields |
+| `observation.state` | **290 float32 values:** joint and end-effector state, including position, orientation, velocity, torque/force, and auxiliary fields |
+| RGB observations | Head, left-wrist, and right-wrist images; video features use CHW metadata |
+| Identity and time | `episode_index`, `task_index`, `frame_index`, and per-step `timestamp` in seconds |
 
-```bash
-pip install "lerobot==0.4.2" "av==15.1.0"
-```
+**End-effector action targets are computed by forward kinematics from joint actions, not independent teleoperation pose commands.** End-effector positions use meters in `base_link`; orientations use XYZW quaternions. Gripper values are not normalized. Select the representation your model expects; do not assume that all 52 action values are joint angles.
 
-```python
-from pathlib import Path
+<details>
+<summary><strong>Action reference: all 52 dimensions</strong></summary>
 
-from huggingface_hub import HfApi, snapshot_download
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+Python slices are stop-exclusive.
 
-repo_id = "wzlllong/value-model-benchmark-training-data"
-prefix = "data/real-world/full-shot"
-revision = HfApi(token=False).dataset_info(repo_id).sha
-snapshot = snapshot_download(
-    repo_id=repo_id,
-    repo_type="dataset",
-    revision=revision,
-    allow_patterns=[f"{prefix}/**"],
-    token=False,
-)
-dataset = LeRobotDataset(
-    repo_id=repo_id,
-    root=Path(snapshot) / prefix,
-    revision=revision,
-    video_backend="pyav",
-)
-sample = dataset[0]
-print(dataset.num_episodes, dataset.num_frames)
-print(sample["action"].shape, sample["observation.state"].shape)
-```
+| Slice | Meaning |
+| --- | --- |
+| `[0:2]` | Left/right effector target opening or stroke |
+| `[2:8]` | Arm end-effector target XYZ positions, m, `base_link` |
+| `[8:16]` | Arm end-effector target XYZW orientations, `base_link` |
+| `[16:30]` | Arm joint target positions, rad |
+| `[30:33]` | Head joint target positions, rad |
+| `[33:38]` | Waist joint target positions, rad |
+| `[38:44]` | Base target XYZ linear and XYZ angular velocity |
+| `[44:48]` | Left reference-pose XYZW quaternion |
+| `[48:52]` | Right reference-pose XYZW quaternion |
+
+The last two reference orientations are distinct from the end-effector target orientations. Fields with `dimensions: 0` and `indices: []` do not occupy the vector. Full definitions are in [`meta/info.json`](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/full-shot/meta/info.json).
+
+</details>
+
+<details>
+<summary><strong>State reference: key fields in the 290-dimensional vector</strong></summary>
+
+| Slice | Meaning |
+| --- | --- |
+| `[0:2]` | Left/right effector state |
+| `[2:26]` | End-effector force/torque fields |
+| `[26:32]` | Arm end-effector XYZ positions, m, `base_link` |
+| `[32:56]` | End-effector linear/angular velocity fields |
+| `[56:64]` | Arm end-effector XYZW orientations, `base_link` |
+| `[108:122]` | Arm joint positions, rad |
+| `[122:136]` | Arm joint torques, Nm |
+| `[136:150]` | Arm joint velocities, rad/s |
+| `[164:176]` | Head joint positions, velocities, efforts, and modes |
+| `[176:196]` | Waist joint positions, velocities, efforts, and modes |
+| `[196:203]` | Robot position and orientation in `odom` |
+
+This is a key-field summary, not the entire vector. Additional pose, mode, operator, camera-extrinsic, and reference fields are defined in [`meta/info.json`](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/full-shot/meta/info.json).
+
+</details>
+
+### 2.3 One-Shot
+
+**Use for one-reference conditioning or task-specific adaptation.** The setting selects one episode for each of the same **20 real-world tasks**, for **20 episodes** in total. These are selected from Full-Shot, not an additional task collection.
+
+The [One-Shot directory](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/one-shot) is reserved; its trajectory files and episode selection are not yet released. Until then, the [Full-Shot quick start](#5-download-and-reading) is the available entry point. Choosing your own single episode does not reproduce the official One-Shot selection.
 
-Expected: **2,000 episodes**, **1,530,987 time steps**, `action` shape **[52]**, and `observation.state` shape **[290]**. `dataset[0]` returns one time step, not an entire episode; LeRobot resolves its camera frames from the video metadata.
+## 3. Simulation
+
+### 3.1 Overview
+
+The simulation domain covers **15 manipulation tasks**. Choose it for value-model experiments and demonstration conditioning in simulated environments. It follows the same Full-Shot/One-Shot organization and LeRobot Dataset v3.0 release specification, but its robot, cameras, and state/action schema must be read from its own subset metadata rather than inferred from G2A.
+
+| Setting | Per task | Episodes | Data |
+| --- | ---: | ---: | --- |
+| [Full-Shot](#32-full-shot) | 100 | 1,500 | [Reserved](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/full-shot) |
+| [One-Shot](#33-one-shot) | 1 | 15 | [Reserved](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/one-shot) |
+
+### 3.2 Full-Shot
 
-The example pins the same repository revision for download and reading, and downloads only the chosen subset. A **subset directory** is the LeRobot `root`, not the whole repository. Reserved directories cannot be loaded yet.
+**15 tasks × 100 demonstrations = 1,500 episodes.** Intended for training or fine-tuning with the complete simulation demonstration pool. The [directory](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/full-shot) is reserved; trajectory statistics and the task/schema catalog will accompany the data release.
 
-## 5. Format and Repository
+### 3.3 One-Shot
 
-### 5.1 Shared LeRobot v3.0 Structure
+**15 tasks × 1 demonstration = 15 episodes.** Intended for single-reference conditioning or task-specific adaptation, using one selection per task from simulation Full-Shot. The [directory](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/one-shot) is reserved. There is no simulation subset to load yet.
 
-Each released subset is a self-contained dataset: **per-step Parquet data + camera videos + task and episode metadata**. An episode is one complete task demonstration; v3.0 may store multiple episodes in a chunk file, so **one file does not necessarily mean one episode**.
+## 4. Repository Layout
 
 ```text
 README.md                         English dataset card
 README_zh.md                      Chinese dataset card
 subset_catalog.json               Machine-readable subset index
-assets/                           Dataset-card icons and preview only
 data/
 ├── real-world/
 │   ├── full-shot/                Available LeRobot dataset root
@@ -169,7 +199,7 @@ data/
     └── one-shot/                 Reserved
 ```
 
-The real-world Full-Shot training subset is named **`real-world-full-shot-data`** in [`subset_catalog.json`](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/subset_catalog.json), and lives at `data/real-world/full-shot`. Preview assets are not extra training episodes. Reserved directories contain only a README.
+**A released subset directory is the LeRobot `root`; the repository root is not.** The reserved directories currently contain only a README.
 
 <details>
 <summary><strong>Format reference: files inside the real-world Full-Shot subset</strong></summary>
@@ -191,73 +221,45 @@ Camera feature keys are `observation.images.cam_high`, `observation.images.cam_l
 
 </details>
 
-### 5.2 Real-World Full-Shot: Signals and Statistics
-
-These field definitions apply to **G2A real-world Full-Shot**, not automatically to future simulation data or other robots. Consult each released subset's `meta/info.json` for its own schema.
-
-| Signal / statistic | Representation |
-| --- | --- |
-| Sampling | **30 FPS**, **1,530,987** time steps |
-| Duration | **≈14.18 hours** total; **≈25.52 seconds** per episode on average |
-| RGB observations | **3 views**: head **640 × 400**; left and right wrist **640 × 528** each; CHW image metadata |
-| `action` | **float32[52]**: joint targets **and** end-effector target poses, plus gripper, head, waist, and base-control fields |
-| `observation.state` | **float32[290]**: joint and end-effector state, including position, orientation, velocity, torque/force, and auxiliary fields |
-
-**End-effector action targets are computed by forward kinematics from joint actions, not independent Cartesian teleoperation commands.** End-effector positions use meters in `base_link`; orientations use XYZW quaternions. Gripper values are not normalized. Do not treat all 52 action values as joint angles.
-
-Each step includes `episode_index`, `task_index`, `frame_index`, and `timestamp` in seconds. Episode metadata provides length, data-row boundaries, and per-camera video offsets. Duration and time steps count each trajectory once, not once per camera.
-
-<details>
-<summary><strong>Action reference: all 52 dimensions</strong></summary>
-
-Python slices are stop-exclusive.
-
-| Slice | Meaning |
-| --- | --- |
-| `[0:2]` | Left/right effector target opening or stroke |
-| `[2:8]` | Arm end-effector target XYZ positions, m, `base_link` |
-| `[8:16]` | Arm end-effector target XYZW orientations, `base_link` |
-| `[16:30]` | Arm joint target positions, rad |
-| `[30:33]` | Head joint target positions, rad |
-| `[33:38]` | Waist joint target positions, rad |
-| `[38:44]` | Base target XYZ linear and XYZ angular velocity |
-| `[44:48]` | Left reference-pose XYZW quaternion |
-| `[48:52]` | Right reference-pose XYZW quaternion |
-
-The last two reference orientations are distinct from the end-effector target orientations. Fields with `dimensions: 0` and `indices: []` do not occupy the vector. Full definitions are in [`meta/info.json`](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/full-shot/meta/info.json).
+## 5. Download and Reading
 
-</details>
+**Quick start — currently available Real-World Full-Shot.** Public downloads require no token. The full subset is approximately **23.23 GB**, including videos.
 
-<details>
-<summary><strong>State reference: key fields in the 290-dimensional vector</strong></summary>
+Use Python **3.10+** and a v3.0-compatible reader. The example is verified with **LeRobot 0.4.2** and **PyAV 15.1.0**; **Dataset v3.0 is the data format, not a package named LeRobot 3.0**.
 
-| Slice | Meaning |
-| --- | --- |
-| `[0:2]` | Left/right effector state |
-| `[2:26]` | End-effector force/torque fields |
-| `[26:32]` | Arm end-effector XYZ positions, m, `base_link` |
-| `[32:56]` | End-effector linear/angular velocity fields |
-| `[56:64]` | Arm end-effector XYZW orientations, `base_link` |
-| `[108:122]` | Arm joint positions, rad |
-| `[122:136]` | Arm joint torques, Nm |
-| `[136:150]` | Arm joint velocities, rad/s |
-| `[164:176]` | Head joint positions, velocities, efforts, and modes |
-| `[176:196]` | Waist joint positions, velocities, efforts, and modes |
-| `[196:203]` | Robot position and orientation in `odom` |
+```bash
+pip install "lerobot==0.4.2" "av==15.1.0"
+```
 
-This is a key-field summary, not the entire vector. Additional pose, mode, operator, camera-extrinsic, and reference fields are defined in [`meta/info.json`](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/full-shot/meta/info.json).
+```python
+from pathlib import Path
 
-</details>
+from huggingface_hub import HfApi, snapshot_download
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
 
-## 6. Collection and Selection
+repo_id = "wzlllong/value-model-benchmark-training-data"
+prefix = "data/real-world/full-shot"
+revision = HfApi(token=False).dataset_info(repo_id).sha
+snapshot = snapshot_download(
+    repo_id=repo_id,
+    repo_type="dataset",
+    revision=revision,
+    allow_patterns=[f"{prefix}/**"],
+    token=False,
+)
+dataset = LeRobotDataset(
+    repo_id=repo_id,
+    root=Path(snapshot) / prefix,
+    revision=revision,
+    video_backend="pyav",
+)
+sample = dataset[0]
+print(dataset.num_episodes, dataset.num_frames)
+print(sample["action"].shape, sample["observation.state"].shape)
+```
 
-- **Training pool:** standard in-domain expert demonstrations, **100 episodes per task**. The released real-world split is `train: 0:2000`; diagnostic and OOD evaluation trajectories are not included.
-- **One-Shot unit:** **one complete episode per task**, not per object, scene, or camera. The real-world and simulation selections come from their respective Full-Shot pools and are not independent additional demonstrations.
-- **Official selection:** the selected `episode_index` values and Full-Shot correspondence have not been released. Choosing your own episode is valid for a custom experiment, but does not reproduce the official One-Shot setting.
-- **Supervision:** there are **no extra scalar reward or value-target labels**. Your method must define its supervision or target construction.
+Expected: **2,000 episodes**, **1,530,987 time steps**, `action` shape **[52]**, and `observation.state` shape **[290]**. `dataset[0]` reads one time step, not an entire episode. LeRobot resolves the corresponding camera frames from video metadata.
 
-## 7. License and Citation
+The example pins one repository revision for both download and reading, and downloads only the chosen subset. Do not replace the prefix with a reserved directory: there are no trajectory files to load there yet.
 
 **License:** no data license has been specified. Confirm applicable terms before use or redistribution.
-
-For research context and the paper, see the [RoboValue project page](https://robovalue-benchmark.github.io/): *RoboValue: A Fine-Grained Sim-and-Real Benchmark for Unified Evaluation of Robotic Value Models* (Liu et al.). When using the data, reference the [dataset repository](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data) and the repository revision used in your experiment.

README_zh.md

@@ -2,44 +2,43 @@
 
 **简体中文** · [English](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data)
 
-本仓库面向**机器人价值模型（value model）**，提供**真实世界**与**仿真世界**操作任务的专家示范。每个领域并列提供两种示范设置：**Full-Shot** 用于训练或微调，**One-Shot** 用于示例参考或任务适配。
+本仓库面向**机器人价值模型（value model）**，提供真实世界与仿真操作任务的专家示例。每个领域均分为两种示例使用设置：**Full-Shot** 用于完整示例训练或微调，**One-Shot** 用于单条任务参考或单示例适配。
 
-共用 **LeRobot Dataset v3.0** 发布格式。Full-Shot 为**每任务 100 条轨迹**；One-Shot 为**每任务一条轨迹**，从对应 Full-Shot 中选取，不是额外采集的一组数据。
+各子集统一以 **LeRobot Dataset v3.0** 作为发布格式：逐时间步 Parquet 数据、相机视频，以及任务和轨迹元数据。不同子集的机器人、相机、状态与动作字段以各自说明为准，不要求跨平台完全相同。
 
-[演示预览](#1-演示预览) · [真实世界](#2-真实世界) · [仿真世界](#3-仿真世界) · [快速开始](#4-快速开始) · [数据格式](#5-数据格式与目录) · [选取规则](#6-采集与选取规则)
+**当前可用：真机 Full-Shot，20 个任务，共 2,000 条轨迹。** 其他三类已预留目录并定义规模，目前尚无轨迹文件。
 
-## 1. 演示预览
+[真实世界](#2-真实世界) · [仿真世界](#3-仿真世界) · [快速开始](#5-下载与读取) · [仓库目录](#4-仓库目录)
 
-<video controls muted playsinline preload="none" width="640" height="400" poster="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/resolve/main/assets/previews/real-world-task-00.jpg">
-  <source src="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/resolve/main/assets/previews/real-world-task-00.mp4" type="video/mp4">
-</video>
+## 1. 数据总览
 
-**用托盘转移积木** · 任务 0 · 轨迹 0 · 头部相机 · 23.17 秒。[打开视频](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/resolve/main/assets/previews/real-world-task-00.mp4)。
+| 领域 | 任务数 | Full-Shot | One-Shot |
+| --- | ---: | ---: | ---: |
+| [真实世界](#2-真实世界) | 20 | **2,000** | 20<sup>†</sup> |
+| [仿真世界](#3-仿真世界) | 15 | 1,500<sup>†</sup> | 15<sup>†</sup> |
 
-预览来自已发布真机训练集中的一条完整轨迹。仿真预览将随对应数据提供；这里不使用合成或无关的演示代替。
+表中数量均为**轨迹条数**，不是视频文件数或相机视角数。Full-Shot **每任务 100 条**；One-Shot **从对应 Full-Shot 集合中每任务选择 1 条**，不是额外采集的数据。† 表示预留子集的定义规模。
+
+**什么时候使用哪一种设置？**
+
+- **Full-Shot：**使用完整专家示例集合，训练或微调任务条件下的价值、进度等表征。
+- **One-Shot：**以一条任务示例作为模型参考条件，或在单示例预算下进行任务适配。
+
+本仓库是**标准域内训练示例集**，不包含论文中的诊断轨迹或分布外（OOD）评测轨迹。数据**不额外提供标量奖励或价值目标标签**，训练监督与目标构造需由所用方法自行定义。
 
 ## 2. 真实世界
 
-<img src="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/resolve/main/assets/domain-real-world.svg" width="24" height="24" alt=""> **G2A · 真机操作示范**
+### 2.1 总览
 
-**20 个操作任务**，覆盖物体放置与分类、家居整理、抽屉操作、重复按压、厨房操作及包装。需要基于真实视觉观测和物理交互训练价值模型或提供示例参考时，选择这一领域。
+真实世界数据由 **G2A 机器人**采集，包含 **20 个操作任务**，覆盖物体放置与分类、家居整理、抽屉操作、重复按压、厨房操作及包装。需要基于真实视觉观测和物理交互进行价值模型训练或示例参考时，选择这一领域。
 
-<table width="100%">
-  <thead>
-    <tr><th scope="col" width="20%">示范设置</th><th scope="col" width="40%">Full-Shot</th><th scope="col" width="40%">One-Shot</th></tr>
-  </thead>
-  <tbody>
-    <tr><th scope="row">任务数</th><td>20</td><td>20</td></tr>
-    <tr><th scope="row">轨迹数</th><td><strong>2,000</strong><br>每任务 100 条</td><td><strong>20</strong><br>每任务 1 条</td></tr>
-    <tr><th scope="row">下载大小</th><td><strong>≈23.23 GB</strong></td><td>待公布</td></tr>
-    <tr><th scope="row">适用需求</th><td>训练／微调</td><td>示例参考／任务适配</td></tr>
-    <tr><th scope="row">可用状态</th><td><strong>已发布</strong></td><td>预留</td></tr>
-    <tr><th scope="row">数据入口</th><td><a href="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/full-shot">浏览文件</a> · <a href="#4-快速开始">快速开始</a></td><td><a href="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/one-shot">子集目录</a> · <a href="#6-采集与选取规则">选取规则</a></td></tr>
-  </tbody>
-</table>
+| 设置 | 每任务 | 总轨迹数 | 数据 |
+| --- | ---: | ---: | --- |
+| [**Full-Shot**](#22-full-shot) | 100 | **2,000** | [文件](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/full-shot) |
+| [One-Shot](#23-one-shot) | 1 | 20 | [预留](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/one-shot) |
 
 <details>
-<summary><strong>真机任务清单：20 个任务及简要说明</strong></summary>
+<summary><strong>任务清单：20 个任务及简要说明</strong></summary>
 
 索引对应训练数据中的 `task_index`。名称与简述采用 RoboValue 论文附录 D 的任务定义；`meta/tasks.parquet` 中的原始任务字符串保持不变。
 
@@ -68,91 +67,122 @@
 
 </details>
 
-## 3. 仿真世界
+### 2.2 Full-Shot
 
-<img src="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/resolve/main/assets/domain-simulation.svg" width="24" height="24" alt=""> **仿真操作 · 15 个任务**
+**用于每任务 100 条完整示例的训练或微调。** 这是当前可用的真机训练子集。
 
-适用于仿真环境中的价值模型实验与示例参考。两种设置遵循相同的 LeRobot v3.0 发布规范；机器人、相机、状态与动作字段将按此领域单独说明，不能直接套用 G2A 的定义。
+[下载与读取](#5-下载与读取) · [浏览文件](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/full-shot) · [字段定义](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/full-shot/meta/info.json)
 
-<table width="100%">
-  <thead>
-    <tr><th scope="col" width="20%">示范设置</th><th scope="col" width="40%">Full-Shot</th><th scope="col" width="40%">One-Shot</th></tr>
-  </thead>
-  <tbody>
-    <tr><th scope="row">任务数</th><td>15</td><td>15</td></tr>
-    <tr><th scope="row">轨迹数</th><td><strong>1,500</strong><br>每任务 100 条</td><td><strong>15</strong><br>每任务 1 条</td></tr>
-    <tr><th scope="row">下载大小</th><td>待公布</td><td>待公布</td></tr>
-    <tr><th scope="row">适用需求</th><td>训练／微调</td><td>示例参考／任务适配</td></tr>
-    <tr><th scope="row">可用状态</th><td>预留</td><td>预留</td></tr>
-    <tr><th scope="row">数据入口</th><td><a href="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/full-shot">浏览文件</a></td><td><a href="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/one-shot">子集目录</a> · <a href="#6-采集与选取规则">选取规则</a></td></tr>
-  </tbody>
-</table>
+**关键统计**
 
-**如何理解数量：**当前只有真机 Full-Shot 含轨迹文件；预留子集的数量表示定义规模，不代表已经可下载。One-Shot 复用 Full-Shot 轨迹，统计独立示范数时不能重复相加；未发布子集的下载大小不作推算。
+| 项目 | 数值 |
+| --- | --- |
+| 规模 | **20 个任务 × 100 条 = 2,000 条轨迹** |
+| 时间步与采样率 | **1,530,987** 个时间步，**30 FPS** |
+| 轨迹时长 | 总计约 **14.18 小时**；平均每条约 **25.52 秒** |
+| 相机 | **3 路 RGB：**头部、左腕、右腕 |
+| 分辨率 | 头部 **640 × 400**；每个腕部 **640 × 528** |
+| 存储 | 约 **23.23 GB**；Parquet + HEVC 视频 |
+| 划分 | `train: 0:2000` |
 
+时长与时间步按轨迹统计，不因三路相机重复计数。
 
-## 4. 快速开始
+**一条轨迹包含什么？**
 
-**读取当前可用的真机 Full-Shot。** 公开下载无需 Token；完整子集含视频约 **23.23 GB**。
+一条 **episode** 对应一条完整的任务示例，包含任务与轨迹标识、逐时间步的机器人观测和动作，以及三路相机视频。轨迹元数据记录长度、数据行边界与视频片段位置。v3.0 允许多条轨迹共用分块文件，因此不能简单把一个文件当作一条轨迹。
 
-使用 Python **3.10+**；读取示例已用 **LeRobot 0.4.2** 和 **PyAV 15.1.0** 验证。**Dataset v3.0 是数据格式版本，不是名为 LeRobot 3.0 的软件包版本。**
+| 信号 | 表示 |
+| --- | --- |
+| `action` | **52 维 float32：**同时包含**关节目标与末端目标位姿**，另含夹爪、头部、腰部和底盘控制字段 |
+| `observation.state` | **290 维 float32：**关节与末端状态，包括位置、朝向、速度、力矩／力及辅助字段 |
+| RGB 观测 | 头部、左腕、右腕图像；视频字段的图像元数据使用 CHW 顺序 |
+| 标识与时间 | `episode_index`、`task_index`、`frame_index`，以及单位为秒的逐帧 `timestamp` |
 
-```bash
-pip install "lerobot==0.4.2" "av==15.1.0"
-```
+**末端动作目标由关节 action 通过正运动学（FK）计算，不是另一套独立的遥操作末端指令。** 末端位置以 `base_link` 为坐标系，单位 m；朝向为 XYZW 四元数。夹爪数值未归一化。使用时应选择模型所需的表示，不能把全部 52 维动作都当成关节角度。
 
-```python
-from pathlib import Path
+<details>
+<summary><strong>动作参考：52 维完整索引</strong></summary>
 
-from huggingface_hub import HfApi, snapshot_download
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
+以下为 Python 切片，右端不包含在内。
 
-repo_id = "wzlllong/value-model-benchmark-training-data"
-prefix = "data/real-world/full-shot"
-revision = HfApi(token=False).dataset_info(repo_id).sha
-snapshot = snapshot_download(
-    repo_id=repo_id,
-    repo_type="dataset",
-    revision=revision,
-    allow_patterns=[f"{prefix}/**"],
-    token=False,
-)
-dataset = LeRobotDataset(
-    repo_id=repo_id,
-    root=Path(snapshot) / prefix,
-    revision=revision,
-    video_backend="pyav",
-)
-sample = dataset[0]
-print(dataset.num_episodes, dataset.num_frames)
-print(sample["action"].shape, sample["observation.state"].shape)
-```
+| 切片 | 含义 |
+| --- | --- |
+| `[0:2]` | 左右执行器目标开度或行程 |
+| `[2:8]` | 双臂末端目标 XYZ 位置，m，`base_link` |
+| `[8:16]` | 双臂末端目标 XYZW 朝向，`base_link` |
+| `[16:30]` | 双臂关节目标角度，rad |
+| `[30:33]` | 头部关节目标角度，rad |
+| `[33:38]` | 腰部关节目标角度，rad |
+| `[38:44]` | 底盘目标 XYZ 线速度与 XYZ 角速度 |
+| `[44:48]` | 左参考姿态 XYZW 四元数 |
+| `[48:52]` | 右参考姿态 XYZW 四元数 |
+
+最后两项参考朝向不同于末端目标朝向。`dimensions: 0` 且 `indices: []` 的字段不占用向量。完整定义见 [`meta/info.json`](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/full-shot/meta/info.json)。
+
+</details>
+
+<details>
+<summary><strong>状态参考：290 维向量的主要字段</strong></summary>
+
+| 切片 | 含义 |
+| --- | --- |
+| `[0:2]` | 左右执行器状态 |
+| `[2:26]` | 末端力与力矩字段 |
+| `[26:32]` | 双臂末端 XYZ 位置，m，`base_link` |
+| `[32:56]` | 末端线速度与角速度字段 |
+| `[56:64]` | 双臂末端 XYZW 朝向，`base_link` |
+| `[108:122]` | 双臂关节角度，rad |
+| `[122:136]` | 双臂关节力矩，Nm |
+| `[136:150]` | 双臂关节速度，rad/s |
+| `[164:176]` | 头部关节位置、速度、力矩与模式 |
+| `[176:196]` | 腰部关节位置、速度、力矩与模式 |
+| `[196:203]` | 机器人在 `odom` 下的位置与朝向 |
+
+这里是主要字段摘要，并非全部 290 维。其他位姿、模式、操作员、相机外参及参考字段，以 [`meta/info.json`](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/full-shot/meta/info.json) 为准。
+
+</details>
+
+### 2.3 One-Shot
+
+**用于单条示例参考或任务适配。** 从相同的 **20 个真机任务**中，每任务选择一条 Full-Shot 轨迹，共 **20 条**，不是另一套任务或额外采集的数据。
+
+[One-Shot 目录](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/one-shot)目前为预留位置，轨迹文件与正式示例选择尚未发布。当前可用入口为 [Full-Shot 快速开始](#5-下载与读取)；自行挑选一条轨迹不等于复现正式 One-Shot 的示例选择。
 
-预期：**2,000 条轨迹**、**1,530,987 个时间步**，`action` 维度 **[52]**，`observation.state` 维度 **[290]**。`dataset[0]` 返回一个时间步，不是整条轨迹；LeRobot 根据视频元数据读取对应相机帧。
+## 3. 仿真世界
+
+### 3.1 总览
+
+仿真领域包含 **15 个操作任务**，适用于仿真环境中的价值模型实验与示例参考。组织方式同样为 Full-Shot 与 One-Shot，发布格式同样为 LeRobot Dataset v3.0；具体机器人、相机、状态与动作字段需查看该子集的元数据，不能直接套用 G2A 的字段定义。
+
+| 设置 | 每任务 | 总轨迹数 | 数据 |
+| --- | ---: | ---: | --- |
+| [Full-Shot](#32-full-shot) | 100 | 1,500 | [预留](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/full-shot) |
+| [One-Shot](#33-one-shot) | 1 | 15 | [预留](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/one-shot) |
+
+### 3.2 Full-Shot
 
-示例将下载与读取固定到同一仓库版本，只下载所选子集。LeRobot 的 `root` 应为**子集目录**，不是整个仓库；预留目录目前还不能加载。
+**15 个任务 × 100 条 = 1,500 条轨迹。** 用于完整仿真示例集合的训练或微调。[目录](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/full-shot)已预留，轨迹统计、任务��单与字段说明将随数据发布提供。
 
-## 5. 数据格式与目录
+### 3.3 One-Shot
 
-### 5.1 共用的 LeRobot v3.0 结构
+**15 个任务 × 1 条 = 15 条轨迹。** 从仿真 Full-Shot 中每任务选择一条，用于单示例参考或任务适配。[目录](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/one-shot)已预留，目前尚无可读取的仿真子集。
 
-每个已发布子集都是完整的数据根目录：**逐时间步 Parquet + 相机视频 + 任务与轨迹元数据**。一条 episode 对应一条完整任务示范；v3.0 可以将多条轨迹合并到分块文件，因此**一个文件不一定是一条轨迹**。
+## 4. 仓库目录
 
 ```text
 README.md                         英文数据集卡片
 README_zh.md                      中文数据集卡片
 subset_catalog.json               机器可读子集索引
-assets/                           仅供数据卡展示的图标与预览
 data/
 ├── real-world/
-│   ├── full-shot/                已发布的 LeRobot 数据根目录
+│   ├── full-shot/                当前可用的 LeRobot 数据根目录
 │   └── one-shot/                 预留
 └── simulation/
     ├── full-shot/                预留
     └── one-shot/                 预留
 ```
 
-真机 Full-Shot 训练子集在 [`subset_catalog.json`](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/subset_catalog.json) 中命名为 **`real-world-full-shot-data`**，目录为 `data/real-world/full-shot`。页面预览不计作新增训练轨迹；预留目录只含 README。
+**已发布子集的目录才是 LeRobot 的 `root`，不是整个仓库根目录。** 预留目录目前只包含 README。
 
 <details>
 <summary><strong>格式参考：真机 Full-Shot 子集内部的文件</strong></summary>
@@ -174,73 +204,45 @@ data/
 
 </details>
 
-### 5.2 真机 Full-Shot：信号与统计
-
-以下定义适用于 **G2A 真机 Full-Shot**，不直接代表后续仿真数据或其他机器人。各已发布子集的具体字段以其 `meta/info.json` 为准。
-
-| 信号／统计 | 表示 |
-| --- | --- |
-| 采样 | **30 FPS**，**1,530,987** 个时间步 |
-| 时长 | 总计约 **14.18 小时**；平均每条约 **25.52 秒** |
-| RGB 观测 | **3 路**：头部 **640 × 400**；左右腕各 **640 × 528**；图像元数据为 CHW 顺序 |
-| `action` | **float32[52]**：同时包含**关节目标与末端目标位姿**，另含夹爪、头部、腰部与底盘控制字段 |
-| `observation.state` | **float32[290]**：关节与末端状态，包括位置、朝向、速度、力矩／力及辅助字段 |
-
-**末端动作目标由关节 action 通过正运动学（FK）计算，不是另一套独立的末端遥操作指令。** 末端位置使用 `base_link` 坐标系与米（m）单位；朝向为 XYZW 四元数。夹爪数值未归一化，不能把全部 52 维动作当作关节角度。
-
-每个时间步包含 `episode_index`、`task_index`、`frame_index` 及以秒为单位的 `timestamp`。轨迹元数据记录长度、数据行边界与各相机视频偏移。时长和时间步按轨迹统计，不因三路相机重复计数。
-
-<details>
-<summary><strong>动作参考：52 维完整索引</strong></summary>
-
-以下为 Python 切片，右端不包含在内。
-
-| 切片 | 含义 |
-| --- | --- |
-| `[0:2]` | 左右执行器目标开度或行程 |
-| `[2:8]` | 双臂末端目标 XYZ 位置，m，`base_link` |
-| `[8:16]` | 双臂末端目标 XYZW 朝向，`base_link` |
-| `[16:30]` | 双臂关节目标角度，rad |
-| `[30:33]` | 头部关节目标角度，rad |
-| `[33:38]` | 腰部关节目标角度，rad |
-| `[38:44]` | 底盘目标 XYZ 线速度与 XYZ 角速度 |
-| `[44:48]` | 左参考姿态 XYZW 四元数 |
-| `[48:52]` | 右参考姿态 XYZW 四元数 |
-
-最后两项参考朝向不同于末端目标朝向。`dimensions: 0` 且 `indices: []` 的字段不占用向量。完整定义见 [`meta/info.json`](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/full-shot/meta/info.json)。
+## 5. 下载与读取
 
-</details>
+**快速开始：当前可用的真机 Full-Shot。** 公开下载无需 Token，含视频的完整子集约 **23.23 GB**。
 
-<details>
-<summary><strong>状态参考：290 维向量的主要字段</strong></summary>
+使用 Python **3.10+** 与支持 v3.0 的读取器。示例已通过 **LeRobot 0.4.2** 和 **PyAV 15.1.0** 验证；**Dataset v3.0 是数据格式版本，不是名为 LeRobot 3.0 的软件包版本**。
 
-| 切片 | 含义 |
-| --- | --- |
-| `[0:2]` | 左右执行器状态 |
-| `[2:26]` | 末端力与力矩字段 |
-| `[26:32]` | 双臂末端 XYZ 位置，m，`base_link` |
-| `[32:56]` | 末端线速度与角速度字段 |
-| `[56:64]` | 双臂末端 XYZW 朝向，`base_link` |
-| `[108:122]` | 双臂关节角度，rad |
-| `[122:136]` | 双臂关节力矩，Nm |
-| `[136:150]` | 双臂关节速度，rad/s |
-| `[164:176]` | 头部关节位置、速度、力矩与模式 |
-| `[176:196]` | 腰部关节位置、速度、力矩与模式 |
-| `[196:203]` | 机器人在 `odom` 下的位置与朝向 |
+```bash
+pip install "lerobot==0.4.2" "av==15.1.0"
+```
 
-这里是主要字段摘要，并非全部 290 维。其他位姿、模式、操作员、相机外参及参考字段，以 [`meta/info.json`](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/full-shot/meta/info.json) 为准。
+```python
+from pathlib import Path
 
-</details>
+from huggingface_hub import HfApi, snapshot_download
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
 
-## 6. 采集与选取规则
+repo_id = "wzlllong/value-model-benchmark-training-data"
+prefix = "data/real-world/full-shot"
+revision = HfApi(token=False).dataset_info(repo_id).sha
+snapshot = snapshot_download(
+    repo_id=repo_id,
+    repo_type="dataset",
+    revision=revision,
+    allow_patterns=[f"{prefix}/**"],
+    token=False,
+)
+dataset = LeRobotDataset(
+    repo_id=repo_id,
+    root=Path(snapshot) / prefix,
+    revision=revision,
+    video_backend="pyav",
+)
+sample = dataset[0]
+print(dataset.num_episodes, dataset.num_frames)
+print(sample["action"].shape, sample["observation.state"].shape)
+```
 
-- **训练池：**标准域内专家示范，**每任务 100 条**。已发布真机划分为 `train: 0:2000`；不含诊断轨迹或分布外（OOD）评测轨迹。
-- **One-Shot 的一条：每个任务一条完整轨迹**，不是按物体、场景或相机计算。真机与仿真分别从各自 Full-Shot 中选取，不是独立新增的数据。
-- **正式选取：**所选 `episode_index` 及其 Full-Shot 对应关系尚未发布。可以自行选取用于自定义实验，但不等于复现正式 One-Shot 设置。
-- **监督信号：**数据**不额外提供标量奖励或价值目标标签**，训练监督与目标构造由所用方法自行定义。
+预期为 **2,000 条轨迹**、**1,530,987 个时间步**，`action` 维度为 `float32[52]`，`observation.state` 维度为 `float32[290]`。`dataset[0]` 读取的是一个时间步，而不是整条轨迹；LeRobot 根据视频元数据读取对应的相机帧。
 
-## 7. 许可证与引用
+示例固定同一个仓库版本用于下载和读取，并且只下载指定子集。不要直接把路径换成预留目录：这些目录目前尚无可加载的轨迹文件。
 
 **许可：**数据许可尚未指定，使用或再分发前请确认适用条款。
-
-研究背景与论文见 [RoboValue 项目主页](https://robovalue-benchmark.github.io/)：*RoboValue: A Fine-Grained Sim-and-Real Benchmark for Unified Evaluation of Robotic Value Models*（Liu 等）。使用数据时，请注明[数据集仓库](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data)及实验采用的仓库版本。