Query-oriented Micro-video Summarization (QMS)

Dataset and trained checkpoint for Jill0001/QMS. QMS generates short, search-query-style summaries of micro-videos using visual and textual inputs.

Dataset preview

The Dataset Viewer above displays all 60,096 records, with the original 17 columns and original split names:

Split	Records
train	50,096
eval	5,000
test	5,000

The Parquet files are lossless conversions of qms_dataset.ft, partitioned by its type column. Each split was read back and compared with its source Arrow table. Within each split, the original row order is retained.

Example records

Excerpts from the original records (long text is shortened here; the Parquet files retain the complete values).

Video ID	Category	Transcript excerpt	Target query example
31984481115	动物	今天发现BB我不见了，哪里都没有BB昨天晚上笼子盖没有关，应该是跑出去了，但是这个笼子很高，一般跑不出去，加油找b区b区吧。	仓鼠不见了怎么办
33375845680	才艺	所以住万三另外的小路看我像吗，我每天都要做这么像漫步又会忘了身处妖精跳出物神秘的心情，要对全世界说，所以住万三另外在小路看我像吗？我和都要做这么像漫步又会忘了。	适合新手画的动漫人物
42203361643	健康,生活	管的这些小技巧，知道你就捡到宝了，一插入吸管时容易弯折，只需用大拇指封住上端，就可以轻松他好二用剪刀给吸管，这样剪开剜着一下，就能把下水道的头发轻松取出来，三用小刀把西瓜呈螺旋状还看用它来收纳家里的电…	生活小技巧小妙招

Downloads

File	Contents	Bytes
final_clip.zip	Original visual feature archive	2,747,572,218
qms_dataset.ft	Original Feather table with transcripts, queries and metadata	183,580,722
saved_checkpoint.pth	Original trained PyTorch checkpoint	740,597,057

The three original files are preserved byte for byte. File hashes are listed in SHA256SUMS.

Fields

Column	Description
item_id	Video identifier, stored as a string
keyword	Original query/keyword string
kw_less	Comma-separated target queries used by the QMS training loader
caption	Video caption
category_name	Content category
asr_pure	Speech-recognition transcript used by the training loader
asr_result	Serialized ASR output, including segment information
ocr_title, ocr_cover, ocr_result	OCR text and serialized OCR output
query_nouns	Query noun annotations
cap_in_q, asr_in_q	Caption/transcript terms associated with queries
query_sim	Query similarity scores used for target sampling
type	Original split: train, eval, or test
path, frame_paths	Original local feature and frame paths

The path columns retain the original dataset values and refer to the author's filesystem. Adapt feature paths to your extraction directory when training. The downloads contain visual features; raw video and image files are not included in this release.

Visual feature format

final_clip.zip contains 60,096 feature files under final_clip/, named <item_id>_clip.npy. These IDs match every record in the text table. The original files use pickled NumPy arrays despite the .npy extension; the source loader reads them with np.load(path, allow_pickle=True). The archive also includes original macOS metadata entries.

Usage

from datasets import load_dataset

dataset = load_dataset("JillJia/QMS")
print(dataset["train"][0]["asr_pure"])
print(dataset["train"][0]["kw_less"])

Download an original artifact:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="JillJia/QMS",
    repo_type="dataset",
    filename="final_clip.zip",
)

See the source repository for model code and training instructions.