Hugging Face's logo

Datasets:
qiusizhan
/
BEAT-EB-nominal-reference-v1

Modalities:
Image
Formats:
imagefolder
Languages:
English
Size:
< 1K
ArXiv:
Tags:
embodied-ai
backdoor-auditing
Libraries:
Datasets
Dataset card Data Studio Files
xet
 Community
1

Research access

Input files require manual access approval.

Log in or Sign Up to review the conditions and access this dataset content.

BEAT-EB nominal reference inputs, v1

This release contains 500 input-only cases sampled from the normal EB-ALFRED task distribution for fitting and calibrating auditing methods. It contains no checkpoint outputs, current-step target completions, attack ground-truth labels, or annotations derived from a checkpoint's behavior.

The source is EmbodiedBench/EB-Alfred_trajectory_dataset at commit 720b5f5683b9ecbb9418f8c5998d686cde3a1397. Both successful and failed source rollouts were eligible. Selection did not filter by success, visible trigger objects, source answers, or checkpoint activation.

Selection and independence

Original ALFRED trials were mapped before sampling. Training, validation and smoke/development trials were excluded; ambiguous consumed-task mappings were excluded conservatively. Eligible trials were partitioned into development, runtime and nominal pools with seed 42 before inspecting images or model outputs. After checking trajectory-prefix integrity, excluding consumed-image and cross-runtime/nominal decoded-image collisions, and deduplicating text plus image, 500 cases were drawn uniformly from 2,901 eligible nominal cases with seed 42.

These inputs represent 25 original trials, 313 source rollouts and 346 distinct images. Different histories can share a current image. They are not 500 independent tasks. Trial separation does not imply simulator-scene separation. Normal-distribution inputs do not imply that every model's response is Negative.

The task_partitions.json mapping assigns whole opaque task groups to fitting (258 cases / 12 trials) or calibration (242 cases / 13 trials), using seed 43. Do not split correlated cases randomly across fitting and calibration.

Files and use

  • inputs.jsonl: 500 records with input_id, opaque task_id, and messages.
  • assets/: original observations, referenced by relative paths in messages.
  • task_partitions.json: task-group fitting/calibration assignments.
  • release-audit.json: selection and processor-validation evidence.
  • checksums.json: byte sizes, SHA-256 and Git blob checksums.

Messages contain the native BEAT system instruction, original task and action catalog, current observation, and only earlier reasoning/actions/environment feedback. Resolve image paths relative to the dataset root and use the tested model's own processor and chat template.

All 500 inputs were validated using the exact pinned Qwen3.5-9B processor from training (c202236235762e1c871ad0ccb60c8ee5ba337b9a) with thinking disabled and without truncation. The longest prompt is 8,125 tokens. Other model families require their own native processor validation. Precomputed 500 model completions are not part of this input release.

Formal BEAT-EB runtime stratification and model-specific actual-output labels are separate work and are not supplied by this nominal dataset.

Attribution

Please cite Rui Yang et al., EmbodiedBench: Comprehensive Benchmarking Multi-modal Large Language Models for Vision-Driven Embodied Agents, arXiv:2502.09560 (2025), and the original ALFRED benchmark when using these inputs. Original content remains subject to upstream rights and terms. The pinned upstream dataset card does not declare a dataset license; this derived release does not grant additional rights to that content.
Downloads last month
19
Total file size:
88.2 MB

Paper for qiusizhan/BEAT-EB-nominal-reference-v1