Datasets:

wzlllong
/

value-model-benchmark-training-data

# RoboValue Small-Scale Simulation Public Test Set

[简体中文](README_zh.md) · [Repository overview](../../../README.md) · [Simulation subsets](../README.md) · [All 30 trajectories](TRAJECTORIES.md) · [Parquet catalog](episode_catalog.parquet)

This **In-Domain diagnostic test subset** contains **30 complete simulation trajectories across four tasks**. It provides RGB video inputs, English instructions, and separate annotations for **10 metric entry points**. Standard demonstrations, alternative orders, error/recovery branches, and complete memory histories are retained. Its task and solution coverage is smaller than the complete RoboValue leaderboard test set.

The subset is stored under `evaluation/simulation/test-30/`, separately from training demonstrations in `training-data/`. Local episode and task indices belong to this subset. Native metadata retains `train: 0:30` for LeRobot container compatibility; the intended use and HF trajectory-catalog split are **test**.

## Tasks and trajectories

| Local task index | Task | Episodes | Local episode indices |
| ---: | --- | ---: | --- |
| 0 | Fill Pen Holder | 12 | 0–11 |
| 1 | Hang Mugs | 12 | 12–23 |
| 2 | Press By Number | 3 | 24–26 |
| 3 | Swap Blocks | 3 | 27–29 |

The subset contains **17,149 synchronized time steps at 25 FPS** and **90 AV1 videos**, approximately **0.595 GB (decimal)**. Every episode has one front view and two wrist views, all **640 × 480**, with no audio. Time steps and duration are counted once per episode, not once per camera. Full episodes and their causal history are preserved.

Trajectory types are **ED4 / MS8 / ER4 / IR4 / EC4 / MEM6**. Stored terminal-success labels comprise **22 true and 8 false**. These label counts describe supplied annotations: ED/MS labels and semantic alignment are user-attested, while the 12 TRR branches and 6 MEM episodes retain the status **pending independent terminal evidence**. Technical video/loader checks do not upgrade these statuses to independently certified terminal outcomes. The evidence is preserved per episode and in [`annotation_provenance.json`](annotations/annotation_provenance.json).

`task_index` 0–3 and `episode_index` 0–29 are public local indices. They are independent of training and real-world test subset indices. Source episode identifiers, collection dates, and acquisition paths are absent.

<a id="metric-coverage"></a>

## Which trajectories evaluate which metrics?

[`annotations/metric_membership.json`](annotations/metric_membership.json) is the authoritative membership file. The same asset can participate in several metrics: **86 metric-member references correspond to 30 stored episodes**.

Definitions follow the paper’s §4.3 (pp. 9–12): **higher is better except FPL (lower is better)**. The paper’s Memory-VOC uses the dataset label VOC-MEM and key `voc_mem`. See the [complete metric summary](../../../README.md#metric-summary).

| Metric / code key | Brief definition | Eligible trajectories | Episodes / groups | Supporting annotations |
| --- | --- | --- | --- | --- |
| TGA-CT / `tga_easy` | Measures how often the correct instruction gives a larger start-to-end value gain than every cross-task negative instruction for the same successful episode. | Pen-holder and mug ED | 4 | `tasks.json`, `task_candidates.json` (13 original candidates) |
| TGA-CF / `tga_hard` | Measures how often the correct instruction gives a larger start-to-end value gain than every counterfactual instruction changing the object, action, placement, or constraint. | Pen-holder and mug ED | 4 | `tasks.json` (4 counterfactual instructions per regular task) |
| VOC / `voc` | Measures the Spearman rank correlation between predicted values and normalized time along successful trajectories. | Pen-holder and mug ED | 4 | Original subtask intervals and sampling anchors in `episodes.jsonl` |
| Cycle-VOC / `cycle_voc` | Averages Spearman correlations with forward and reversed progress order when a successful trajectory is replayed forward then backward as one continuous sequence. | Pen-holder and mug ED | 4 | Complete episodes, original intervals, and sampling anchors |
| VS / `vs` | Combines trend efficiency with non-flat time coverage to reward stable value updates while penalizing reversals and prolonged flat intervals. | Pen-holder and mug ED | 4 | Same forward episodes and sampling basis used by Cycle-VOC |
| SA / `sa` | Measures the fraction of same-task successful/unsuccessful episode pairs in which the successful episode has the higher average terminal score. | All pen-holder and mug trajectories | 24; 64 same-task pairs | Stored labels and evidence states in `episodes.jsonl`, original pairings in `sa_pairs.jsonl` |
| SIA / `sia` | Scores subtask recognition as the geometric mean of the independent judge’s probabilities for the correct subtask, based on the model’s free-form descriptions. | Pen-holder and mug ED/MS | 12 | Original subtask descriptions and intervals |
| TRR / `trr` | Checks matched trajectories for the expected stage trends: failure down, continued error non-increasing, recovery attempt up, and successful/failed recovery outcome up/down. | Pen-holder and mug ER/IR/EC | 12 / 4 three-role groups | `trr_groups.json`, original failure/recovery stages in `episodes.jsonl` |
| CSVC / `csvc` | Measures the consistency of local value gains for semantically matched subtasks across valid execution solutions. | Pen-holder and mug ED/MS | 12 / 6 order groups | `csvc_solutions.json`, independent `semantic_alignment.json` |
| VOC-MEM / `voc_mem` | Applies VOC’s Spearman score to successful long trajectories with recurring visual states, using the history available to each model interface. | Press-by-number and swap-blocks MEM | 6 | Complete videos and original `voc_mem` sampling anchors |

The four ED episodes are `0, 1, 12, 13`; MEM episodes are `24–29`. The catalog's `metrics` column is generated from this membership file. Counts across metrics overlap and must not be summed to obtain the number of trajectories.

| Trajectory type | Supplied role | Count | Applicable metrics |
| --- | --- | ---: | --- |
| ED | Standard successful demonstration | 4 | TGA-CT, TGA-CF, VOC, Cycle-VOC, VS, SA, SIA, CSVC |
| MS | Successful alternative order | 8 | SA, SIA, CSVC |
| ER | Effective recovery branch | 4 | SA, TRR |
| IR | Ineffective recovery branch | 4 | SA, TRR |
| EC | Error-continuation branch | 4 | SA, TRR |
| MEM | Complete memory-history episode | 6 | VOC-MEM |

Original FPL fields remain inside per-episode annotations where supplied. This subset has **no formal FPL cohort or FPL metric entry**. TRR groups preserve their members and stages; no unannotated target subtask is inferred. MEM task subtask tables and per-episode `subtask_segments` are **empty as supplied**: sampling anchors and complete history are available, but counting-subtask intervals or per-frame progress ground truth have not been added.

## Chronological and cross-order semantic annotations

SIA uses the original chronological subtask numbering and intervals. CSVC additionally uses stable object/action identities recorded in [`semantic_alignment.json`](annotations/semantic_alignment.json). Both annotation layers are retained; cross-order semantic alignment does not replace or renumber the chronological labels.

Each regular task has **three order groups**, with **two executions per group**: the standard ED order and two alternative MS orders. Together they form **6 groups and 12 episodes**. Their grouping, color order, semantic action order, and per-instance intervals are preserved in [`csvc_solutions.json`](annotations/csvc_solutions.json). This is partial order coverage, not exhaustive coverage of all legal solutions. Supplied evidence for success, independent execution, object identity, and legal order remains user attestation.

Left/right roles are preserved by episode. Standard pen-holder executions use right-arm insertion with left-arm holding; alternative executions use left-arm insertion with right-arm holding. Mug color/action identities and recorded arm assignments are retained. Arms and video views are not mirrored or normalized across order groups.

## Files and annotation semantics

```text
meta/info.json                         Native format, features, FPS, path templates
meta/tasks.parquet                     English instructions indexed by task_index
meta/episodes/chunk-000/file-000.parquet Episode lengths and camera/video references
data/chunk-000/file-*.parquet           Native per-frame location records
videos/observation.images.front/       Front-view videos
videos/observation.images.wrist_left/  Left-wrist videos
videos/observation.images.wrist_right/ Right-wrist videos
annotations/                          Separate evaluation annotations
episode_catalog.parquet               One browsing row per episode
TRAJECTORIES.md                        Human-readable trajectory and video index
checksums.sha256                       Release-file integrity manifest
```

Native data records contain only `timestamp`, `frame_index`, `episode_index`, `index`, and `task_index`. Original float32 episode-relative timestamps and local frame indices are retained; public episode/task indices and the cumulative global `index` are rebuilt. These fields locate images and instructions, rather than providing model observations.

`codebase_version: v3.0` identifies the native format, and `robot_type` is null. State/action vectors, old robot feature metadata, collection information, and source identifiers are omitted. Numeric normalization statistics are omitted because the export has no numeric observation features. Container/encoder additions are cleared with stream copying; frames are not re-encoded, resized, cropped, resampled, or reordered.

| Annotation file | Contents |
| --- | --- |
| [`tasks.json`](annotations/tasks.json) | Four local task indices, original English instructions, chronological subtask descriptions, and counterfactual instructions |
| [`episodes.jsonl`](annotations/episodes.jsonl) | Thirty local episode/task indices, trajectory roles, stored success labels and evidence states, complete original annotation objects |
| [`metric_membership.json`](annotations/metric_membership.json) | Exact local episode/task cohorts for the 10 metric entry points |
| [`task_candidates.json`](annotations/task_candidates.json) | Complete 13-instruction cross-task pool and candidate correspondence |
| [`trr_groups.json`](annotations/trr_groups.json) | Four local groups, ER/IR/EC members, branch roles, and stage intervals |
| [`csvc_solutions.json`](annotations/csvc_solutions.json) | Six order groups, two executions per group, color/semantic order, and instance relationships |
| [`semantic_alignment.json`](annotations/semantic_alignment.json) | Stable object/action identities, original chronological IDs, semantic segments, and left/right roles |
| [`sa_pairs.jsonl`](annotations/sa_pairs.jsonl) | Original 64 same-task success/failure pairings mapped to local indices |
| [`annotation_provenance.json`](annotations/annotation_provenance.json) | Supplied attestation and verification-evidence meaning, with local episode references |

All frame indices are local and zero-based. Intervals retain their original endpoint values. `start_frame` / `end_frame_exclusive` ranges and legacy FPL `failure_start` / `failure_end` ranges are interpreted as **half-open `[start, end)`**, without adding one to the end. Existing event and action endpoint conventions are retained.

The `annotations` object in each episode preserves the original annotation content. Source-reference fields are mapped to public local episode/group indices. Sampling `frame_index` entries are anchors, not ground-truth value/progress scores or model predictions. Semantic IDs name object/action concepts, and group/solution indices express evaluation relationships rather than acquisition identities.

## Trajectory catalog and HF preview

The `simulation-test-30` configuration explicitly exposes a **`test` split with 30 episode rows** from `episode_catalog.parquet`. Per-frame samples are read with LeRobot. Auxiliary reports and navigation documents are excluded from preview discovery. [`TRAJECTORIES.md`](TRAJECTORIES.md) lists every episode's eligible metrics and direct links to all three camera videos.

Both public test domains use the same 12 catalog columns:

| Field | Type | Meaning |
| --- | --- | --- |
| `episode_index`, `task_index` | `int64` | Indices local to this subset |
| `task_name`, `instruction` | `string` | English display task name and original instruction |
| `trajectory_type` | `string` | ED, MS, ER, IR, EC, or MEM |
| `success` | `bool` | Supplied terminal-success label; read its evidence state separately |
| `num_frames` | `int64` | Episode time-step count, counted once across views |
| `duration_s` | `float64` | Duration (`num_frames / fps`) |
| `video_primary`, `video_left_wrist`, `video_right_wrist` | `string` | Paths relative to this subset root; primary is front for simulation, head for real-world |
| `metrics` | `list[string]` | Eligible code keys from the formal membership file |

Video paths and labels support browsing and evaluation. They are not decoded model observations. Stored success-label evidence is available through the matching entry in `annotations/episodes.jsonl` and the provenance file.

## Download and load

The example is pinned to `578593ef28ee7a2d861ebf38162c928291b0d4a9`, which contains both public test sets, and downloads only this subset. Download and load using the same fixed revision.

```python
from pathlib import Path
from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset
repo_id = "wzlllong/value-model-benchmark-training-data"
prefix = "evaluation/simulation/test-30"
revision = "578593ef28ee7a2d861ebf38162c928291b0d4a9"
snapshot = snapshot_download(
    repo_id=repo_id,
    repo_type="dataset",
    revision=revision,
    allow_patterns=[f"{prefix}/**"],
    local_dir="./robovalue-data",
)
root = Path(snapshot) / prefix
dataset = LeRobotDataset(
    repo_id=repo_id,
    root=root,
    revision=revision,
    video_backend="pyav",
    tolerance_s=1e-4,
)
item = dataset[0]
value_input = {
    "images": {camera: item[camera] for camera in dataset.meta.camera_keys},
    "instruction": item["task"],
}
print(dataset.num_episodes, dataset.num_frames)  # 30, 17149
```

For an extracted local upload package, set `root = Path("upload/evaluation/simulation/test-30")` and load it with the same constructor, omitting the remote `revision` argument. The root must directly contain `meta/`, `data/`, and `videos/`. Camera keys are `observation.images.front`, `observation.images.wrist_left`, and `observation.images.wrist_right`; loaded images use CHW shape `[3, 480, 640]`.

Send selected images and the instruction to the value model. Read annotation labels, groups, arm roles, trajectory types, stored success flags, and metric membership separately. Native location fields and catalog paths are routing metadata, rather than model observations. History-based models can use `delta_timestamps` to request earlier frames in episode order.

```python
import json
episodes = {
    row["episode_index"]: row
    for row in (
        json.loads(line)
        for line in (root / "annotations/episodes.jsonl").read_text().splitlines()
        if line.strip()
    )
}
episode_index = int(item["episode_index"])
labels = episodes[episode_index]
assert int(item["task_index"]) == labels["task_index"]
membership = json.loads(
    (root / "annotations/metric_membership.json").read_text()
)["metrics"]
eligible_metrics = [
    key for key, cohort in membership.items()
    if episode_index in cohort["episode_indices"]
]
```

From the subset directory, check the release-file hashes:

```bash
shasum -a 256 -c checksums.sha256
```

## Subtask annotations

[`annotations/subtasks.json`](annotations/subtasks.json) provides S-numbered definitions, source citations, and correspondence to original label IDs; [`annotations/subtask_segments.parquet`](annotations/subtask_segments.parquet) retains every existing segment. Intervals use `[start_frame, end_frame_exclusive)`, missing intervals remain absent, and metric cohorts and object/action semantics remain in the original annotations. See [Subtask reading](../../../SUBTASKS.md) for loading examples.

	# RoboValue Small-Scale Simulation Public Test Set

	[简体中文](README_zh.md) · [Repository overview](../../../README.md) · [Simulation subsets](../README.md) · [All 30 trajectories](TRAJECTORIES.md) · [Parquet catalog](episode_catalog.parquet)

	This In-Domain diagnostic test subset contains 30 complete simulation trajectories across four tasks. It provides RGB video inputs, English instructions, and separate annotations for 10 metric entry points. Standard demonstrations, alternative orders, error/recovery branches, and complete memory histories are retained. Its task and solution coverage is smaller than the complete RoboValue leaderboard test set.

	The subset is stored under `evaluation/simulation/test-30/`, separately from training demonstrations in `training-data/`. Local episode and task indices belong to this subset. Native metadata retains `train: 0:30` for LeRobot container compatibility; the intended use and HF trajectory-catalog split are test.

	## Tasks and trajectories

	\| Local task index \| Task \| Episodes \| Local episode indices \|
	\| ---: \| --- \| ---: \| --- \|
	\| 0 \| Fill Pen Holder \| 12 \| 0–11 \|
	\| 1 \| Hang Mugs \| 12 \| 12–23 \|
	\| 2 \| Press By Number \| 3 \| 24–26 \|
	\| 3 \| Swap Blocks \| 3 \| 27–29 \|

	The subset contains 17,149 synchronized time steps at 25 FPS and 90 AV1 videos, approximately 0.595 GB (decimal). Every episode has one front view and two wrist views, all 640 × 480, with no audio. Time steps and duration are counted once per episode, not once per camera. Full episodes and their causal history are preserved.

	Trajectory types are ED4 / MS8 / ER4 / IR4 / EC4 / MEM6. Stored terminal-success labels comprise 22 true and 8 false. These label counts describe supplied annotations: ED/MS labels and semantic alignment are user-attested, while the 12 TRR branches and 6 MEM episodes retain the status pending independent terminal evidence. Technical video/loader checks do not upgrade these statuses to independently certified terminal outcomes. The evidence is preserved per episode and in [`annotation_provenance.json`](annotations/annotation_provenance.json).

	`task_index` 0–3 and `episode_index` 0–29 are public local indices. They are independent of training and real-world test subset indices. Source episode identifiers, collection dates, and acquisition paths are absent.

	<a id="metric-coverage"></a>

	## Which trajectories evaluate which metrics?

	[`annotations/metric_membership.json`](annotations/metric_membership.json) is the authoritative membership file. The same asset can participate in several metrics: 86 metric-member references correspond to 30 stored episodes.

	Definitions follow the paper’s §4.3 (pp. 9–12): higher is better except FPL (lower is better). The paper’s Memory-VOC uses the dataset label VOC-MEM and key `voc_mem`. See the [complete metric summary](../../../README.md#metric-summary).

	\| Metric / code key \| Brief definition \| Eligible trajectories \| Episodes / groups \| Supporting annotations \|
	\| --- \| --- \| --- \| --- \| --- \|
	\| TGA-CT / `tga_easy` \| Measures how often the correct instruction gives a larger start-to-end value gain than every cross-task negative instruction for the same successful episode. \| Pen-holder and mug ED \| 4 \| `tasks.json`, `task_candidates.json` (13 original candidates) \|
	\| TGA-CF / `tga_hard` \| Measures how often the correct instruction gives a larger start-to-end value gain than every counterfactual instruction changing the object, action, placement, or constraint. \| Pen-holder and mug ED \| 4 \| `tasks.json` (4 counterfactual instructions per regular task) \|
	\| VOC / `voc` \| Measures the Spearman rank correlation between predicted values and normalized time along successful trajectories. \| Pen-holder and mug ED \| 4 \| Original subtask intervals and sampling anchors in `episodes.jsonl` \|
	\| Cycle-VOC / `cycle_voc` \| Averages Spearman correlations with forward and reversed progress order when a successful trajectory is replayed forward then backward as one continuous sequence. \| Pen-holder and mug ED \| 4 \| Complete episodes, original intervals, and sampling anchors \|
	\| VS / `vs` \| Combines trend efficiency with non-flat time coverage to reward stable value updates while penalizing reversals and prolonged flat intervals. \| Pen-holder and mug ED \| 4 \| Same forward episodes and sampling basis used by Cycle-VOC \|
	\| SA / `sa` \| Measures the fraction of same-task successful/unsuccessful episode pairs in which the successful episode has the higher average terminal score. \| All pen-holder and mug trajectories \| 24; 64 same-task pairs \| Stored labels and evidence states in `episodes.jsonl`, original pairings in `sa_pairs.jsonl` \|
	\| SIA / `sia` \| Scores subtask recognition as the geometric mean of the independent judge’s probabilities for the correct subtask, based on the model’s free-form descriptions. \| Pen-holder and mug ED/MS \| 12 \| Original subtask descriptions and intervals \|
	\| TRR / `trr` \| Checks matched trajectories for the expected stage trends: failure down, continued error non-increasing, recovery attempt up, and successful/failed recovery outcome up/down. \| Pen-holder and mug ER/IR/EC \| 12 / 4 three-role groups \| `trr_groups.json`, original failure/recovery stages in `episodes.jsonl` \|
	\| CSVC / `csvc` \| Measures the consistency of local value gains for semantically matched subtasks across valid execution solutions. \| Pen-holder and mug ED/MS \| 12 / 6 order groups \| `csvc_solutions.json`, independent `semantic_alignment.json` \|
	\| VOC-MEM / `voc_mem` \| Applies VOC’s Spearman score to successful long trajectories with recurring visual states, using the history available to each model interface. \| Press-by-number and swap-blocks MEM \| 6 \| Complete videos and original `voc_mem` sampling anchors \|

	The four ED episodes are `0, 1, 12, 13`; MEM episodes are `24–29`. The catalog's `metrics` column is generated from this membership file. Counts across metrics overlap and must not be summed to obtain the number of trajectories.

	\| Trajectory type \| Supplied role \| Count \| Applicable metrics \|
	\| --- \| --- \| ---: \| --- \|
	\| ED \| Standard successful demonstration \| 4 \| TGA-CT, TGA-CF, VOC, Cycle-VOC, VS, SA, SIA, CSVC \|
	\| MS \| Successful alternative order \| 8 \| SA, SIA, CSVC \|
	\| ER \| Effective recovery branch \| 4 \| SA, TRR \|
	\| IR \| Ineffective recovery branch \| 4 \| SA, TRR \|
	\| EC \| Error-continuation branch \| 4 \| SA, TRR \|
	\| MEM \| Complete memory-history episode \| 6 \| VOC-MEM \|

	Original FPL fields remain inside per-episode annotations where supplied. This subset has no formal FPL cohort or FPL metric entry. TRR groups preserve their members and stages; no unannotated target subtask is inferred. MEM task subtask tables and per-episode `subtask_segments` are empty as supplied: sampling anchors and complete history are available, but counting-subtask intervals or per-frame progress ground truth have not been added.

	## Chronological and cross-order semantic annotations

	SIA uses the original chronological subtask numbering and intervals. CSVC additionally uses stable object/action identities recorded in [`semantic_alignment.json`](annotations/semantic_alignment.json). Both annotation layers are retained; cross-order semantic alignment does not replace or renumber the chronological labels.

	Each regular task has three order groups, with two executions per group: the standard ED order and two alternative MS orders. Together they form 6 groups and 12 episodes. Their grouping, color order, semantic action order, and per-instance intervals are preserved in [`csvc_solutions.json`](annotations/csvc_solutions.json). This is partial order coverage, not exhaustive coverage of all legal solutions. Supplied evidence for success, independent execution, object identity, and legal order remains user attestation.

	Left/right roles are preserved by episode. Standard pen-holder executions use right-arm insertion with left-arm holding; alternative executions use left-arm insertion with right-arm holding. Mug color/action identities and recorded arm assignments are retained. Arms and video views are not mirrored or normalized across order groups.

	## Files and annotation semantics

	```text
	meta/info.json Native format, features, FPS, path templates
	meta/tasks.parquet English instructions indexed by task_index
	meta/episodes/chunk-000/file-000.parquet Episode lengths and camera/video references
	data/chunk-000/file-*.parquet Native per-frame location records
	videos/observation.images.front/ Front-view videos
	videos/observation.images.wrist_left/ Left-wrist videos
	videos/observation.images.wrist_right/ Right-wrist videos
	annotations/ Separate evaluation annotations
	episode_catalog.parquet One browsing row per episode
	TRAJECTORIES.md Human-readable trajectory and video index
	checksums.sha256 Release-file integrity manifest
	```

	Native data records contain only `timestamp`, `frame_index`, `episode_index`, `index`, and `task_index`. Original float32 episode-relative timestamps and local frame indices are retained; public episode/task indices and the cumulative global `index` are rebuilt. These fields locate images and instructions, rather than providing model observations.

	`codebase_version: v3.0` identifies the native format, and `robot_type` is null. State/action vectors, old robot feature metadata, collection information, and source identifiers are omitted. Numeric normalization statistics are omitted because the export has no numeric observation features. Container/encoder additions are cleared with stream copying; frames are not re-encoded, resized, cropped, resampled, or reordered.

	\| Annotation file \| Contents \|
	\| --- \| --- \|
	\| [`tasks.json`](annotations/tasks.json) \| Four local task indices, original English instructions, chronological subtask descriptions, and counterfactual instructions \|
	\| [`episodes.jsonl`](annotations/episodes.jsonl) \| Thirty local episode/task indices, trajectory roles, stored success labels and evidence states, complete original annotation objects \|
	\| [`metric_membership.json`](annotations/metric_membership.json) \| Exact local episode/task cohorts for the 10 metric entry points \|
	\| [`task_candidates.json`](annotations/task_candidates.json) \| Complete 13-instruction cross-task pool and candidate correspondence \|
	\| [`trr_groups.json`](annotations/trr_groups.json) \| Four local groups, ER/IR/EC members, branch roles, and stage intervals \|
	\| [`csvc_solutions.json`](annotations/csvc_solutions.json) \| Six order groups, two executions per group, color/semantic order, and instance relationships \|
	\| [`semantic_alignment.json`](annotations/semantic_alignment.json) \| Stable object/action identities, original chronological IDs, semantic segments, and left/right roles \|
	\| [`sa_pairs.jsonl`](annotations/sa_pairs.jsonl) \| Original 64 same-task success/failure pairings mapped to local indices \|
	\| [`annotation_provenance.json`](annotations/annotation_provenance.json) \| Supplied attestation and verification-evidence meaning, with local episode references \|

	All frame indices are local and zero-based. Intervals retain their original endpoint values. `start_frame` / `end_frame_exclusive` ranges and legacy FPL `failure_start` / `failure_end` ranges are interpreted as half-open `[start, end)`, without adding one to the end. Existing event and action endpoint conventions are retained.

	The `annotations` object in each episode preserves the original annotation content. Source-reference fields are mapped to public local episode/group indices. Sampling `frame_index` entries are anchors, not ground-truth value/progress scores or model predictions. Semantic IDs name object/action concepts, and group/solution indices express evaluation relationships rather than acquisition identities.

	## Trajectory catalog and HF preview

	The `simulation-test-30` configuration explicitly exposes a `test` split with 30 episode rows from `episode_catalog.parquet`. Per-frame samples are read with LeRobot. Auxiliary reports and navigation documents are excluded from preview discovery. [`TRAJECTORIES.md`](TRAJECTORIES.md) lists every episode's eligible metrics and direct links to all three camera videos.

	Both public test domains use the same 12 catalog columns:

	\| Field \| Type \| Meaning \|
	\| --- \| --- \| --- \|
	\| `episode_index`, `task_index` \| `int64` \| Indices local to this subset \|
	\| `task_name`, `instruction` \| `string` \| English display task name and original instruction \|
	\| `trajectory_type` \| `string` \| ED, MS, ER, IR, EC, or MEM \|
	\| `success` \| `bool` \| Supplied terminal-success label; read its evidence state separately \|
	\| `num_frames` \| `int64` \| Episode time-step count, counted once across views \|
	\| `duration_s` \| `float64` \| Duration (`num_frames / fps`) \|
	\| `video_primary`, `video_left_wrist`, `video_right_wrist` \| `string` \| Paths relative to this subset root; primary is front for simulation, head for real-world \|
	\| `metrics` \| `list[string]` \| Eligible code keys from the formal membership file \|

	Video paths and labels support browsing and evaluation. They are not decoded model observations. Stored success-label evidence is available through the matching entry in `annotations/episodes.jsonl` and the provenance file.

	## Download and load

	The example is pinned to `578593ef28ee7a2d861ebf38162c928291b0d4a9`, which contains both public test sets, and downloads only this subset. Download and load using the same fixed revision.

	```python
	from pathlib import Path
	from huggingface_hub import snapshot_download
	from lerobot.datasets.lerobot_dataset import LeRobotDataset

	repo_id = "wzlllong/value-model-benchmark-training-data"
	prefix = "evaluation/simulation/test-30"
	revision = "578593ef28ee7a2d861ebf38162c928291b0d4a9"
	snapshot = snapshot_download(
	repo_id=repo_id,
	repo_type="dataset",
	revision=revision,
	allow_patterns=[f"{prefix}/**"],
	local_dir="./robovalue-data",
	)
	root = Path(snapshot) / prefix
	dataset = LeRobotDataset(
	repo_id=repo_id,
	root=root,
	revision=revision,
	video_backend="pyav",
	tolerance_s=1e-4,
	)
	item = dataset[0]
	value_input = {
	"images": {camera: item[camera] for camera in dataset.meta.camera_keys},
	"instruction": item["task"],
	}
	print(dataset.num_episodes, dataset.num_frames) # 30, 17149
	```

	For an extracted local upload package, set `root = Path("upload/evaluation/simulation/test-30")` and load it with the same constructor, omitting the remote `revision` argument. The root must directly contain `meta/`, `data/`, and `videos/`. Camera keys are `observation.images.front`, `observation.images.wrist_left`, and `observation.images.wrist_right`; loaded images use CHW shape `[3, 480, 640]`.

	Send selected images and the instruction to the value model. Read annotation labels, groups, arm roles, trajectory types, stored success flags, and metric membership separately. Native location fields and catalog paths are routing metadata, rather than model observations. History-based models can use `delta_timestamps` to request earlier frames in episode order.

	```python
	import json

	episodes = {
	row["episode_index"]: row
	for row in (
	json.loads(line)
	for line in (root / "annotations/episodes.jsonl").read_text().splitlines()
	if line.strip()
	)
	}
	episode_index = int(item["episode_index"])
	labels = episodes[episode_index]
	assert int(item["task_index"]) == labels["task_index"]
	membership = json.loads(
	(root / "annotations/metric_membership.json").read_text()
	)["metrics"]
	eligible_metrics = [
	key for key, cohort in membership.items()
	if episode_index in cohort["episode_indices"]
	]
	```

	From the subset directory, check the release-file hashes:

	```bash
	shasum -a 256 -c checksums.sha256
	```

	## Subtask annotations

	[`annotations/subtasks.json`](annotations/subtasks.json) provides S-numbered definitions, source citations, and correspondence to original label IDs; [`annotations/subtask_segments.parquet`](annotations/subtask_segments.parquet) retains every existing segment. Intervals use `[start_frame, end_frame_exclusive)`, missing intervals remain absent, and metric cohorts and object/action semantics remain in the original annotations. See [Subtask reading](../../../SUBTASKS.md) for loading examples.