DRIVE-UGC

Decision-stage Ranking with Image and Verbal Evidence from User-Generated Content is a Chinese automotive benchmark for ranking 16 candidate vehicles from collections of user posts and images. It contains 15,850 events, 89,274 posts, and 105,241 selected-image references. The image archive contains 105,043 unique images, identified by content hashes.

The release provides research-processed text, selected original image bytes, event and content-group metadata, and frozen post and image-region features. It supports the cross-stage ranking task used by CAMPER, including bottleneck group-wise feature purification (BGFP) and relative region purification (RRP).

Source and stage definitions

The research collection was assembled for a 2026 study from publicly accessible Autohome automotive-community contributions. An event groups a user's retained posts and associated selected images with one recorded target-vehicle label.

Stages follow observable evidence of target-vehicle choice or ownership. Target events contain eligible posts preceding the selected evidence post. Source events contain retained choice, ownership, or use evidence when eligible target input is absent. These operational stages describe the evidence available in the collection.

Partitions

Partition	Role	All events	Paired events	Posts	Image references
B_train	Source training	9,080	8,786	18,794	59,854
A_train	Target training	4,189	3,343	37,789	25,469
A_val	Target validation	823	823	11,554	6,367
A_test	Target test	1,758	1,758	21,137	13,551
Total		15,850	14,710	89,274	105,241

The complete benchmark includes 1,140 events without images. The matched multimodal experiments use the paired subset, selected by the metadata field paired. Validation and test are entirely paired. Filtering to paired events preserves each event's original public index and feature offsets.

Event identifiers, content-group identifiers, labels, and partition assignments retain their archived associations. Content groups link related records; retain these groups when sampling training subsets. labels.json maps integer labels 0–15 to the 16 Chinese vehicle names.

Released text, images, and features

Text. The records contain the archived, sanitized title and body of each retained post. Vehicle, brand, and trim expressions are represented by the research preprocessing placeholders. Release preparation additionally replaced 14 contact strings across nine posts with [CONTACT]: 11 explicit mobile contact numbers, one QQ identifier, and two phone-number-formatted usernames; the frozen features retain their original encoding, so these nine released texts differ from the encoded texts by those replacements. Posts retain their archived order within each event, aligned with the event's text-feature rows.

Images. TAR shards contain the selected original image bytes from the research archive. The image index maps each content hash to its shard and member path. Event records reference these images by hash and path; they do not embed image bytes.

Text features. BAAI/bge-large-zh-v1.5 supplies one L2-normalized, 1,024-dimensional post vector, stored as float32. Each post is encoded from its sanitized title and body with a 512-token limit.

Visual features. google/siglip2-so400m-patch14-384 supplies the OCR-MASK regional features used by the current BGFP/RRP method. Each image contributes 81 normalized region vectors arranged as a 9×9 grid, with 1,152 coordinates stored as float16. These come from averaging non-overlapping 3×3 neighborhoods of the encoder's 27×27 patch grid.

The selected original images and OCR-MASK features are distinct release components: the image archive preserves selected source bytes, while the feature arrays reflect the benchmark's detected vehicle-identity-text masking procedure. The arrays retain region structure for RRP aggregation. Encoder weights are obtained from their upstream repositories.

Vehicle-identity masking targets textual cues for vehicle ranking. Selected original images may contain recognizable people, license plates, logos, and other identifying details. Public event IDs describe the metadata representation and do not establish complete anonymity of source content. See SOURCE_NOTICE.md for provenance and rights.

Repository layout

README.md
SOURCE_NOTICE.md
MANIFEST.json
labels.json
statistics.json
records/
  B_train.jsonl
  A_train.jsonl
  A_val.jsonl
  A_test.jsonl
splits/{split}/
  events.csv
  events.npz
features/{split}/
  text.npy
  text_event_offsets.npy
  regions.npy
  region_event_offsets.npy
  region_image_offsets.npy
images/
  image_index.jsonl
  images-00000.tar
  images-00001.tar
  ...
code/
  load_features.py

The four split names are B_train, A_train, A_val, and A_test. Image shards are approximately 512 MiB each, with a smaller final shard. Feature arrays occupy approximately 20 GB in total. Download only the partitions and image shards needed for an analysis. MANIFEST.json provides file sizes and SHA-256 hashes for integrity verification.

Metadata and offsets

events.csv contains event_id, content_group_id, label, paired, post_count, and image_count, in public event order. events.npz stores the aligned arrays ids, groups, y, and paired.

For event index i, text_event_offsets[i:i+2] gives its half-open post-row interval in text.npy. region_event_offsets[i:i+2] gives its region-row interval in regions.npy. region_image_offsets.npy preserves image boundaries within the concatenated region array. Events with no images have an empty region interval.

Each JSONL record contains event_id, group_id, label, paired, posts, and images. The record's group_id corresponds to content_group_id in the CSV and feature loader. Each post has sanitized_title and sanitized_text; each image reference has path, sha256, width, height, and post_index. Event and within-event ordering preserve the archived feature alignment.

images/image_index.jsonl contains sha256, path, archive, and bytes for each unique image. Image paths have the form images/<hash-prefix>/<sha256>.<extension> and match the member names inside their TAR archive. The Hub configuration selects only records/{split}.jsonl for the viewer; label mappings, statistics, feature arrays, and TAR files remain separate assets. Image references are paths and hashes rather than decoded image columns. This follows the Hub's explicit data-file configuration.

Load a downloaded partition

Download features/A_test/, splits/A_test/, labels.json, and code/load_features.py into the repository layout above. Run from a directory containing the downloaded repository:

from pathlib import Path
import sys

root = Path("DRIVE-UGC").resolve()
sys.path.insert(0, str(root / "code"))
from load_features import load_split

ds = load_split(root, "A_test")
event = ds.event(int(ds.paired_indices[0]))

print(event["event_id"], event["label"])
print(event["posts"].shape)    # (number of posts, 1024)
print(event["regions"].shape)  # (81 * number of images, 1152)

boundaries = event["image_region_offsets"]
first_image_regions = event["regions"][boundaries[0]:boundaries[1]]

The loader returns public_event_index, event_id, content_group_id, label, paired, posts, regions, and event-relative image_region_offsets. Use the public event index when joining feature slices to metadata; paired-subset positions are a different indexing space.

To inspect local text records:

import json

with (root / "records" / "A_test.jsonl").open(encoding="utf-8") as f:
    record = json.loads(next(f))
print(record.keys())

For original images, resolve the record's image hash through the image index, then read the named member from its TAR shard:

import tarfile

with (root / "images" / "image_index.jsonl").open(encoding="utf-8") as f:
    image_index = {item["sha256"]: item for item in map(json.loads, f)}

image_ref = record["images"][0]  # A_test events all contain images
entry = image_index[image_ref["sha256"]]
with tarfile.open(root / entry["archive"], "r") as archive:
    with archive.extractfile(entry["path"]) as image_file:
        image_bytes = image_file.read()

This retrieves selected source-image bytes. The released regional arrays provide the corresponding OCR-MASK model inputs.

Attribution and rights

Source contributions originate from Autohome. Text features use BGE-large-zh-v1.5, and regional features use SigLIP 2 SO400M patch14-384. Refer to the upstream model cards for encoder documentation and terms.

Third-party posts and images remain subject to the rights of their original rights holders. This repository grants no additional open-content license for those source materials. SOURCE_NOTICE.md describes the released transformations and source rights.