Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Parquet error: Scan size limit exceeded: attempted to read 413374232 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files
Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

miniReranker multimodal training pairs

This dataset supports supervised point-wise multimodal reranking and embedder-guided visual token pruning for miniReranker: Efficient Multimodal Reranking through Visual Cache Reuse and Interaction Sparsity.

Hub repository: Mickey25/miniReranker_736k. Code: EIT-NLP/miniRanker. The release contains 736,327 query-document pairs across 25 subsets and 66 Parquet shards.

Composition

There are 25 subsets and 66 Parquet shards, with one train split. Counts refer to point-wise pairs, not unique queries, images, videos, or documents. No separate validation/test split is supplied. Benchmark evaluation uses external data.

Candidate type Label Pairs
positive yes 147,899
hard_negative no 294,375
random_negative no 294,053
Total 736,327

Each positive was paired with up to four negatives: two hard and two random negatives. According to the paper, hard negatives were mined with Qwen3-VL-Embedding-2B from Top-20 image/document retrieval candidates and Top-5 video retrieval candidates. Actual retained counts are listed below.

Subset/configuration Retrieval direction Pairs Shards
A-OKVQA Image + text → Text 8,525 1
CIRR Image + text → Image 13,055 1
ChartQA Image + text → Text 14,145 1
DocVQA Image + text → Text 19,730 1
HatefulMemes Image → Text 1,700 1
ImageNet_1K Image → Text 50,000 3
InfographicsVQA Image + text → Text 11,970 1
MSCOCO Image → Image 50,000 3
MSCOCO_i2t Image → Text 56,640 3
MSCOCO_t2i Text → Image 50,000 3
N24News Image → Text 24,490 2
NIGHTS Image + text → Image 7,970 1
OK-VQA Image + text → Text 4,500 1
SUN397 Image → Text 9,925 1
ShareGPTVideo_caption_t2v Text → Video 29,971 8
ShareGPTVideo_caption_v2t Video → Text 29,998 8
ShareGPTVideo_qa_v2t Video + text → Text 24,243 6
VOC2007 Image → Text 3,920 1
ViDoRe Text → Document image 59,095 3
VisDial Text → Image + text 61,640 4
VisRAG Text → Document image 61,375 4
Visual7W Image + text → Text 34,905 2
VisualNews_i2t Image → Text 50,000 3
VisualNews_t2i Text → Image + text 49,950 3
WebQA Text → Image + text 8,580 1

Storage and fields

Each subset contains original-*.parquet with the same Arrow schema.

Fields Type Meaning
subset string Source subset identifier, matching its configuration name.
source_row_idx int64 Row index in the source subset; not a globally unique pair ID.
candidate_type string positive, hard_negative, or random_negative.
label string Point-wise relevance target: yes or no.
instruction string Task-specific reranking instruction.
query_text, document_text string Text on either side; may be empty.
query_image_bytes, document_image_bytes binary, nullable Encoded image bytes; no external image download is needed for these fields.
doc_modality string Document-side modality metadata; examples include text, image, image+text, video.
query_video_frame_paths, document_video_frame_paths list[string] Ordered paths to external video frames; empty for non-video sides.
query_visual_modalities, document_visual_modalities list[string] Visual input types, e.g. image or video, in media order on that side.
query_visual_token_counts, document_visual_token_counts list[int32] Number of visual tokens per media input before pruning.
query_visual_ranked_tokens, document_visual_ranked_tokens list[list[int32]] Descending importance ranking as zero-based [media_index, token_index] pairs.

Rankings are computed separately for the query and document. For example, [[0, 12], [0, 3]] means token 12 of media input 0 is more important than token 3. The media_index refers to a visual input on that side, not a query/document row or necessarily an individual video frame. Nonvisual sides have empty ranking lists. The training code selects high-ranked tokens and restores their original sequence order, keeping at least one token for each nonempty media input.

The inspected preprocessing used Qwen3-VL-Embedding-2B attention, summed over all layers from the final embedding token to visual tokens (with head averaging), seed 42, maximum input length 10,240, image budgets 4,096–1,843,200 pixels, video frame budget 786,432 pixels, total video budget 7,864,320 pixels, at most 64 frames, and FPS metadata 2.0. Video rows were retained with probability 0.2 during this preprocessing stage. Do not apply this 0.2 sampling again when training on this release. Training's default video sampling ratio is 1.0. Training/evaluation FPS metadata defaults to 1; retain the stored frame ordering and matching pixel budgets so token counts agree. If geometry differs, recompute rankings rather than applying stale token indices.

Load a subset or the complete collection

Load the published dataset:

from datasets import load_dataset
all_pairs = load_dataset("Mickey25/miniReranker_736k", split="train")
video_pairs = load_dataset("Mickey25/miniReranker_736k",
                           "ShareGPTVideo_caption_t2v", split="train")

Local loading and image decoding:

from datasets import load_dataset
from PIL import Image
from io import BytesIO
pairs = load_dataset("parquet", data_files={"train": "data/miniReranker_736k/VisRAG/original-*.parquet"},
                     split="train")
row = pairs[0]
image = Image.open(BytesIO(row["document_image_bytes"])).convert("RGB")

Download video frames separately

The Parquet files contain paths only, not video files or frame bytes. Download ShareGPTVideo train_300k, which the upstream dataset card describes as the video-frame subset used for SFT/DPO. The three video configurations are caption text-to-video, caption video-to-text, and video question/answer-to-text.

hf download ShareGPTVideo/train_video_and_instruction --repo-type dataset \
  --include 'train_300k/*' --local-dir ./downloads/ShareGPTVideo

Extract the downloaded archives while preserving their relative directory hierarchy. Set the root to the directory containing train_300k, not train_300k itself:

data/ShareGPTVideo/                 <- video_root
└── train_300k/
    └── <video_id>/
        ├── c01_0001.jpeg
        └── c01_0002.jpeg

For a stored path train_300k/<video_id>/c01_0001.jpeg, the loader resolves video_root / stored_path. Configure --video_root ./data/ShareGPTVideo, SHAREGPTVIDEO_ROOT, or VIDEO_ROOT. Missing frames raise an error.

Upload preparation and source terms

Upload only this README and the 25 subset folders' original-*.parquet files. Exclude training_split_cache/, inspection.ipynb, and manifest.worker*.json: the manifests contain machine-specific paths and are not needed to load the dataset. Do not upload ShareGPTVideo frames as part of this repository.

The collection contains derived examples from the named source datasets. Consult the original source terms before redistributing or using each subset; the code's Apache-2.0 license does not replace their licenses. This card does not assign one blanket license to the underlying images, documents or video frames. Dataset names are provenance labels, and a pair count does not describe each original dataset's size.

Relevant upstream collections: MMEB/VLM2Vec, ViDoRe, VisRAG, and ShareGPTVideo. See the code repository's paper for construction and task details.

Downloads last month
374