ActionLens

ActionLens is a video benchmark for evaluating fine-grained, temporally grounded action understanding in vision-language models. Its multiple-choice questions test whether a model can track what happens over time, identify which person performs an action, and interpret concurrent and directed interactions.

Diagnostics

ID	Diagnostic	Examples
D1	Transition Sensitivity	194
D2	Actor-Specific Action Disambiguation	2,000
D3	Concurrent Action Binding	2,000
D4	Directed Interaction Reasoning	1,611
D5	Gaze Direction Attribution	896
Total		6,701

Each example includes a question, four answer options, the answer letter, segment and source-video identifiers, and task-specific metadata. Clips are provided in clips/<diagnostic-folder>/; the video_path field in each JSON record is relative to the repository root.

Repository layout

clips/       Video clips referenced by the annotations
dataset/     Full benchmark annotations, one JSON file per diagnostic
subset/      Smaller evaluation subset and its sampling manifest

The subset contains up to 250 examples per diagnostic and is intended for quick evaluation. Use dataset/ for full benchmark runs.

Load the annotations

from datasets import load_dataset

data = load_dataset(
    "json",
    data_files="hf://datasets/anonymous-dataset-submission/actionlens/dataset/2_actor_disambiguation.json",
    split="train",
)
print(data[0]["question"])

The JSON records store relative video paths. Evaluation code should resolve these paths against the repository or download the referenced clip files from this dataset repository.

Evaluation with lmms-eval

The ActionLens task integration is available on the add-actionlens-dataset branch. From a checkout of that branch, install the project dependencies and run a short smoke evaluation:

python -m lmms_eval \
  --model qwen3_vl \
  --model_args pretrained=Qwen/Qwen3-VL-8B-Instruct \
  --tasks actionlens \
  --limit 5 \
  --batch_size 1 \
  --log_samples \
  --output_path results/actionlens

The integration downloads clips on first use and caches them under the Hugging Face cache directory. Set HF_HOME to choose another cache location. Remove --limit 5 for a full run. Individual diagnostics can be selected with --tasks d1_transition_sensitivity, d2_actor_action_disambiguation, d3_concurrent_action_binding, d4_directed_interaction_reasoning, or d5_gaze_direction_attribution.

Data and license

The dataset repository declares Apache-2.0. The video clips are derived from AVA source videos and include visual annotations such as person boxes; users should also follow the source dataset and video terms when accessing, using, or redistributing the clips.