OneSearch-VL-RL-10K

This dataset contains 10,059 unique training tasks spanning single-image, multi-image and video inputs. Only the training split is released. The schema below lists every field written to the released JSONL files. No unused fields or modality-specific null placeholders are included.

Modality	Train and paired GT rows
Single image	3,691
Multi-image	2,780
Video	3,588
Total	10,059

Files

• train.jsonl: the fixed training split, with one row per task. No test JSONL is published.
• data/{single_image,multi_image,video}.jsonl: the same task schema, grouped by modality and containing only the training tasks; needed to build the current reward index.
• evgr_gt/{single_image,multi_image,video}.jsonl: matching judge-only targets, paired with data/ by row ID and order.
• media_zips/single_image/: 1 ZIP containing 3,691 original images.
• media_zips/multi_image/: 1 ZIP containing 8,336 PNG frames.
• media_zips/video/: 21 ZIPs containing 3,337 prepared MP4s and 426,040 PNG frames.

Each ZIP is independently extractable and smaller than 10 GB. Videos are the prepared 1 fps training inputs, accompanied by all input frames needed by frame selection and Ground reward. Extracted paths start with media/ and match the JSONL references. Frames shared across modalities have identical paths and bytes.

Download and extract

hf download OneSearch-VL/OneSearch-VL-RL-10K --repo-type dataset \
  --local-dir /data/OneSearch-VL-RL-10K
cd /data/OneSearch-VL-RL-10K
for shard in media_zips/single_image/*.zip media_zips/multi_image/*.zip media_zips/video/*.zip; do
    unzip -o "$shard" -d .
done

The media ZIPs total approximately 214.42 GB. Allow additional space for the extracted media; downloaded ZIPs can be removed after successful extraction.

Task JSONL: exact keys and values

These schemas apply to train.jsonl and all three data/ files.

Field	Value	Consumer
id	Unique string task ID	Split integrity, GT pairing and logs
question	Nonempty question string, preserved from the selected data	Actor; judge question context
answer	Nonempty reference answer string, preserved without rewriting	Accuracy and Query; task identity
modality	Exactly single_image, multi_image or video	Input, prompt, tool and Accuracy dispatch
images	Ordered relative-path list; exactly one for single image, at least two for multi-image	Initial image pixels and handles
image_frame_indices	Multi-image only: distinct nonnegative integer frame IDs, one per ordered image	Match each image to Ground targets
video_grid	Video only: an object containing exactly video_url and frame_urls	Video input and frame-selection tools
video_grid.video_url	Relative path to the prepared 1 fps grid MP4	Actual actor video input
video_grid.frame_urls	Nonempty ordered list of distinct relative PNG paths	Frame selection and Ground reference pixels

The single-image keys are id/question/answer/modality/images. The multi-image keys add image_frame_indices. The video keys are id/question/answer/modality/video_grid; there is no images placeholder.

Video frame IDs are the zero-based positions in frame_urls. The number of frames is the list length; grid playback is always 1 fps. n_frames, frame_indices and grid_fps are derived rather than repeated in the file. This playback contract is distinct from the model's visual sampling rate.

Registration computes evgr_media_identity from relative media references before resolving local paths. It is a runtime field, not a release JSONL field. The actor's initial message receives the question and media, not the reference answer or judge targets.

Illustrative task rows (not actual dataset samples):

{"id":"example-si","question":"Which building is shown?","answer":"Example Hall","modality":"single_image","images":["media/example-si/image.jpg"]}
{"id":"example-mi","question":"Which building appears in the images?","answer":"Example Hall","modality":"multi_image","images":["media/example-mi/frame_004.png","media/example-mi/frame_033.png"],"image_frame_indices":[4,33]}
{"id":"example-video","question":"Which two objects appear?","answer":"A bottle and a clock","modality":"video","video_grid":{"video_url":"media/example-video/grid.mp4","frame_urls":["media/example-video/frame_000.png","media/example-video/frame_001.png"]}}

Reward GT JSONL: exact keys and values

Common keys are only id and claim_facts. The ID must match the paired task. The modality comes from the file and task; the answer is stored only in the task.

Field	Value	Consumer
id	String equal to the paired task ID	Prevent row/GT misalignment
claim_facts	Nonempty list of existing fact strings	Trace targets; not retrieved evidence
gt_bboxes	Single-image only: nonempty list of [x1,y1,x2,y2] boxes on a 0–1000 coordinate scale	Ground reference crops
grounding_contract	Single-image fallback only: exactly source_only	Explicit whole-image reference instead of bbox
grounding_annotations	Multi-image/video only: nonempty ordered list of target objects described below	Ground target frames and appearance
grounding_annotations[].frame_index	Required nonnegative integer: a multi-image source frame ID or a video root-grid position	Locate the target reference
grounding_annotations[].object_description	Optional nonempty appearance string; absent when unavailable	Identify the intended visible object

For single image, exactly one of gt_bboxes and grounding_contract is present. There are 3,690 boxed samples and one explicit whole-image fallback. Single-image GT does not add appearance descriptions.

For multi-image/video, each target frame occurs once. Target order is annotation order; ground_frames(gt) derives the target list. There is no duplicate gt_frames and no GT bbox. Missing descriptions are omitted, not invented from answers or facts.

Illustrative GT rows corresponding to the tasks above:

{"id":"example-si","claim_facts":["The building is Example Hall."],"gt_bboxes":[[100,200,700,900]]}
{"id":"example-mi","claim_facts":["The building is Example Hall."],"grounding_annotations":[{"frame_index":4,"object_description":"a red brick building with a clock tower"}]}
{"id":"example-video","claim_facts":["A bottle and a clock appear in different frames."],"grounding_annotations":[{"frame_index":0,"object_description":"a red bottle held in one hand"},{"frame_index":1}]}

The single-image fallback has the form {"id":"example-fallback","claim_facts":["An existing fact."],"grounding_contract":"source_only"}.

Reward boundary

Accuracy and Query use the reference answer from the task. Trace uses claim_facts plus actual retrieval evidence. Ground uses target frames/boxes/appearance plus the actor's actual visual and retrieval evidence. It does not receive the GT answer, claim facts or actor reasoning. The execution evidence ledger is built from each rollout; do not store a prefilled ledger in training JSONL.

The reward remains Format * (0.6 * Accuracy + 0.2 * Query + 0.2 * (Trace + Ground) / 2). Exhausted Trace/Ground scoring returns zero for the failed component with a failure flag; the successful component is retained.

Register and validate

Run from the training code repository with this dataset downloaded and media extracted:

DATA_ROOT=/data/OneSearch-VL-RL-10K \
  bash RL/rllm/vision_deepresearch_async_workflow/data_prepare/register_rl_dataset.sh \
  --media-root /data/OneSearch-VL-RL-10K

PYTHONPATH=RL/rllm python -m vision_deepresearch_async_workflow.reward_preflight \
  --data-root /data/OneSearch-VL-RL-10K

The training entry reads only the train registry entry and sets data.val_files=[]. No validation dataset, dataloader or validation judge is required. Old registered test entries are ignored; train data is not reused as validation.

Questions, answers, fact targets, object descriptions and media pixels are preserved from the selected training data. Dataset preparation is not a new semantic annotation pass.