Preset Smart Refining Operators

Start Node

• Supported file formats: Applicable to all dataset types. However, it specifically provides data format conversion capabilities for "text > single-turn dialogue, single-turn dialogue with persona, multi-turn dialogue, and multi-turn dialogue with persona."
• Note: All dataset formats will be converted into the platform-compatible format after processing at the start node.

  Provides data format conversion for some text data (single-turn dialogue, single-turn dialogue with persona setting, multi-turn dialogue, and multi-turn dialogue with persona setting). The conversion rules are as follows:

  • Platform-compatible datasets can skip conversion and go straight to the refining process.
  • Non-platform-compatible data (like Alpaca or ShareGPT formats) must first be converted to the platform-compatible format at the start node for subsequent processing by data operators.
• Parameter configuration example

  None

• Conversion example

  Dataset format before processing at the input node: platform-compatible, Alpaca, or ShareGPT format.

  Dataset format after processing at the input node: platform-compatible format.

End Node

• Supported file formats: Applicable to all dataset types. However, it specifically provides data format conversion capabilities for "text > single-turn dialogue, single-turn dialogue with persona, multi-turn dialogue, and multi-turn dialogue with persona." You can also choose to output data in different formats.
• Note:

  After smart refining is complete for a dataset, the end node can convert the data format for the specified data type. The conversion rules are as follows:

  • If the dataset input at the start node is in the platform-compatible format, the end node outputs the dataset in the platform-compatible format by default.
  • If the dataset input at the start node is in a non-platform-compatible format (Alpaca or ShareGPT), the end node outputs the dataset in the same non-platform-compatible format by default.
  • The end node can also select a dataset format different from that of the start node as the final output dataset format.
• Parameter configuration example

  None

• Conversion example

  Dataset format before processing at the input node: any format.

  Dataset format after processing at the output node: any format.

Word Document Content Extraction

• Applicable file format: document > docx
• Parameters:

  Type of the content to be extracted: Extracts text from a Word document and retains the titles and body of the original document, but does not retain images, formulas, headers, and footers. Nested tables cannot be extracted.

• Parameter configuration example:

  No parameters need to be set. By default, the contents, title, and body of the original document are retained, and the images, tables, formulas, headers, and footers are not retained.

• Extraction example

  Local import: {"fileName":"JAVA from Beginner to Master.docx","original_path": "Local Import","text":"JAVA is a cross-platform..."}

  OBS import: {"fileName":"JAVA from Beginner to Master.docx","original_path": "nlp_data/word/JAVA from Beginner to Master.docx","text":"JAVA is a cross-platform..."}

  AI Gallery: {"fileName":"JAVA from Beginner to Master.docx","original_path": "Gallery Subscription","text":"JAVA is a cross-platform..."}

CSV Content Extraction

• Applicable dataset types: Text > single-turn Q&A, single-turn Q&A (with persona), and Q&A sorting.
• Parameter description

  Type of content to be extracted: Reads all text content from a CSV file and generates data in JSON format based on the key value of the file content type template.

• Parameter configuration example

  No parameters need to be set.

• Extraction example

  If the extracted CSV content is "Hello, please introduce yourself. I am Pangu model.", the extracted content is {"context":"Hello, please introduce yourself","target":"I am Pangu model."}

PDF Content Extraction

• Applicable dataset type: Document > PDF.
• Parameter description

  Type of content to be extracted: By default, the text, tables, formulas, and titles are retained. You can select the type to be saved. The types that are not selected will be removed.

  Refined content extraction: indicates whether to support image content extraction after layout analysis.

  Available formats for table extraction: The default format is Latex. The table can be converted to the Markdown format.

• Parameter configuration example

  

• Extraction example

  Local import: {"fileName":"JAVA from Beginner to Master.pdf","original_path": "Local Import","text":"JAVA is a cross-platform..."}.

  OBS import: {"fileName":"JAVA from Beginner to Master.pdf","original_path": "nlp_data/pdf/JAVA from Beginner to Master.pdf","text":"JAVA is a cross-platform..."}.

  AI Gallery: {"fileName":"JAVA from Beginner to Master.pdf","original_path": "Gallery Subscription","text":"JAVA is a cross-platform..."}.

• Operator restrictions

  The PDF content extraction process stops after 24 hours if it handles a lot of data. Split the data before running the process.

Personal Data Anonymization

• Applicable dataset type: Text.
• Parameter description

  Type of content to be converted: Anonymizes sensitive personal information in the text, such as mobile numbers, ID cards, email addresses, URLs, license plate numbers in China, IP addresses, MAC addresses, IMEIs, passports, and vehicle identification numbers. By default, all options are selected. You can also select some of them.

• Parameter configuration example

  

• Conversion example

  Before refining: "Data is from www.test.com."

  After refining: "Data is from *******."

Symbol Standardization

• Applicable dataset type: Text.
• Parameter description

  Type of content to be converted: Non-standard symbols in the text can be converted to standard symbols. The non-standard symbols include spaces, DBC symbols, punctuations, and number symbols. By default, all non-standard symbols are selected. The filtering granularity is character.

• Parameter configuration example

  

Custom Regular Expression Replacement

• Applicable dataset type: Text.
• Parameter description

  Type of content to be converted: Uses the customized regular expression to replace the text content if the data items remain unchanged.

• Parameter configuration example

  

• Conversion example

  Before refining: {"text":"This is the main content aeiou in the test aeiou. "}.

  After refining: {"text":"This is the main content 11111 in the test 11111. "}.

Date and Time Format Conversion

• Applicable dataset type: Text.
• Parameter description

  Type of content to be converted: Automatically identifies the date, time, and week, and converts the date, time, and week based on the selected format. The conversion types include date format, time format, and week format. By default, all of them are selected. You can also select some of them.

• Parameter configuration example

  

• Conversion example

  Before refining: {"text":"Today is Monday, March 3, 2025. The rain is heavy in the morning. "}.

  After refining: {"text":"Today is Monday, 2025-03-03 00:00:00. The rain is heavy in the morning. "}.

Advertisement Data Filtering

• Applicable dataset type: Text.
• Parameter description

  Type of content to be filtered: Deletes a sentence that includes advertisement data from the text, based on a filtering granularity of a sentence.

• Parameter configuration example

  

• Filtering example

  Before refining: {"text":"Specific discount! Buy our products and enjoy a discount of up to 50%! Click the link below to avail the discount at https://example.com. Seize this opportunity now and take action! "}.

  After refining: {"text":""}.

Filtering Abnormal Characters

• Applicable dataset type: Text.
• Parameter description

  Type of content to be converted: Searches for abnormal characters carried in each data record in the dataset and replaces the abnormal characters with null values. The data items remain unchanged. Types of abnormal characters include invisible characters, emojis, web page labels, special characters, garbled characters, and special spaces. By default, all types are selected. You can also select some of them.

• Parameter configuration example

  

• Filtering example

  Before refining: {"text":"Test exception. <style></style>Haha. Limited-time offer! ☺"}.

  After refining: {"text":"Test exception. Haha. Limited-time offer!"}.

Custom Regular Expression Filtering

• Applicable dataset type: Text.
• Parameter description

  Type of content to be filtered: Filters content based on a custom regular expression. The filtering granularity can be character (default) or paragraph.

  Regular expression: Enter the regular expression required for custom regular expression filtering.

  Retain Matching Samples: This parameter is displayed when the type of content to be filtered is paragraph. The default value is false.

• Parameter configuration example

  

• Filtering example

  Filtering out the content following "References"

  Before refining: {"text":"This is the body content. References [1] Author 1, Article 1, Journal 1, 2021.[2] Author 2, Article 2, Journal 2, 2022."}.

  After refining: {"text":"This is the body content. "}.

Custom Keyword Filtering

• Applicable dataset type: Text.
• Parameter description

  Type of content to be filtered: The filtering granularity can be character (default), paragraph, or document. The path of the keyword to be deleted supports keyword import from OBS and text input.

• Parameter configuration example

  

• Filtering example

  For example, filter by keyword test.

  Before refining: {"text":"Keyword test. This is a test data record. "}.

  After refining: {"text":"Keyword. This is a test data record. "}.

Filtering of the Incomplete Sentence at the End of a Paragraph

• Applicable dataset type: Text.
• Parameter description

  Type of content to be filtered: Checks whether the content at the end of a paragraph is complete based on the sentence-level filtering granularity, and deletes the content if the content is incomplete.

• Parameter configuration example

  

• Filtering example

  Before refining: "Java is an object-oriented programming language. Use Java. "

  After refining: "Java is an object-oriented programming language."

Sensitive Word Filtering

• Applicable dataset type: Text.
• Parameter description

  Type of content to be filtered: Automatically detects and filters sensitive data such as pornographic, violent, and political content in the text. Sensitive words need to be preset. The filtering granularity can be character (default), paragraph, or document.

• Parameter configuration example

  

• Filtering example

  Before refining: {"text":"prostitute test"}.

  After refining: {"text":"test"}.

Text Length Filtering

• Applicable dataset type: Text.
• Parameter description

  Type of content to be filtered: Retains data within the specified text length. By default, the length of the characters to be reserved ranges from 100 to 1000 characters, which can be modified. The minimum value is 1.

• Parameter configuration example

  

• Filtering example

  Before refining: {"text": "Test length"}

  After refining: {"text":""}

Sentence Feature Filtering

• Applicable dataset type: Text.
• Parameter description

  Type of content to be filtered: Filters the content based on the document filtering granularity and the average sentence length to be retained. If the content does not meet the requirements, the content is filtered out. The default value is greater than or equal to 10 characters, which can be modified. The minimum value is 1.

• Parameter configuration example

  

• Filtering example

  Before refining: {"text":"In a small village, there is a legend. In the legend, a mysterious fox appears in the village forest every full moon night. "}.

  After refining: {"text":""}.

Prohibited Text Detection

• Applicable dataset types: Q&A sorting, single-turn Q&A, and single-turn Q&A (with persona) > JSONL.
• Parameter description: If filtering is enabled, the filtering operator is used. Otherwise, the filtering operator is not used.

• Parameter configuration example

  

• Filtering example

  Before labeling:

  {"text": "Do you have QQ sales shareholder data?"}

  After labeling:

  {"text":"Do you have QQ sales shareholder data?","text_ban_moderation":{"suggestion":"block","details":{"confidence":1.0,"label":"violation_info","risk_level":2,"segments":[{"segment":"QQ sales shareholder data"},{"segment":"Shareholder data"},{"segment":"Shareholder data & sales"},{"segment":"Sales shareholder data"}],"suggestion":"block"}}}

  suggestion indicates whether the file passes the check. pass indicates that the file passes the check and no problem occurs. review indicates that manual review is required. You can choose to bypass or block the file based on your review policy. block indicates that the file to be reviewed is problematic.

Personal Privacy Identification

• Applicable dataset types: Q&A sorting, single-turn Q&A, and single-turn Q&A (with persona) > JSONL.
• Parameter description: If filtering is enabled, the filtering operator is used. Otherwise, the filtering operator is not used.

• Parameter configuration example

  

• Filtering example

  Before labeling:

  {"text": "You save my MAC address: 20-6E-D4-88-F3-98"}

  After labeling:

  {"text": "You save my MAC address: 20-6E-D4-88-F3-98","text_pii_moderation":{"suggestion":"block","details":[{"start":33,"end":50,"length":17,"data":"20-6E-D4-88-F3-98","category":"MAC_ADDRESS"}]}}

  suggestion indicates whether the file passes the check. pass indicates that the file passes the check and no problem occurs. review indicates that manual review is required. You can choose to bypass or block the file based on your review policy. block indicates that the file to be reviewed is problematic.

Garbage Content Text Detection

• Applicable dataset types: Q&A sorting, single-turn Q&A, and single-turn Q&A (with persona) > JSONL.
• Parameter description: If filtering is enabled, the filtering operator is used. Otherwise, the filtering operator is not used.

• Parameter configuration example

  

• Filtering example

  Before labeling:

  {"text": "[Kaiyuan fake certificate 848777596_qq Hefei fake certificate uhc0tm] What does it mean_Kaiyuan false certificate 848777596_qq Hefei false certificate uhc0tm Translation_Phonetic mark_Pronunciation_Usage_Example sentence_Online translation_Youdao dictionary"}

  After labeling:

  {"text":"[Kaiyuan fake certificate 848777596_qq Hefei fake certificate uhc0tm] What does it mean_Kaiyuan false certificate 848777596_qq Hefei false certificate uhc0tm Translation_Phonetic mark_Pronunciation_Usage_Example sentence_Online translation_Youdao dictionary","text_spam_moderation":{"details":[{"confidence":1.0,"label":"abuse","risk_level":2,"segments":[{"segment":"tm's"}]},"suggestion":"block"]},"suggestion":"block"}}

  suggestion indicates whether the file passes the check. pass indicates that the file passes the check and no problem occurs. review indicates that manual review is required. You can choose to bypass or block the file based on your review policy. block indicates that the file to be reviewed is problematic.

Ad Text Detection

• Applicable dataset types: Q&A sorting, single-turn Q&A, and single-turn Q&A (with persona) > JSONL.
• Parameter description: If filtering is enabled, the filtering operator is used. Otherwise, the filtering operator is not used.

• Parameter configuration example

  

• Filtering example

  Before labeling:

  {"context":"On sale for inventory clearance. All items are only CNY2.","target":"The price is cheap."}

  After labeling:

  {"context":"On sale for inventory clearance. All items are only CNY2.","target":"The price is cheap.","text_ad_moderation":{"details":[],"suggestion":"pass"}}

  suggestion indicates whether the file passes the check. pass indicates that the file passes the check and no problem occurs. review indicates that manual review is required. You can choose to bypass or block the file based on your review policy. block indicates that the file to be reviewed is problematic.

Pornographic Text Detection

• Applicable dataset types: Q&A sorting, single-turn Q&A, and single-turn Q&A (with persona) > JSONL.
• Parameter description: If filtering is enabled, the filtering operator is used. Otherwise, the filtering operator is not used.

• Parameter configuration example

  

• Filtering example

  Before labeling:

  {"text": "XXX navigation, come and enjoy hardcore action now, fill your life with erotica and excitement, wait no more..."}

  After labeling:

  {"text":"XXX navigation, come and enjoy hardcore action now, fill your life with erotica and excitement, wait no more...","text_porn_moderation":{"suggestion":"block","details":"[{'confidence': 1.0, 'label': 'porn_violence', 'risk_level': 2, 'segments': [{'segment': 'hardcore action'}, {'segment': 'XXX navigation'}], 'suggestion': 'block'}]"}}

  suggestion indicates whether the file passes the check. pass indicates that the file passes the check and no problem occurs. review indicates that manual review is required. You can choose to bypass or block the file based on your review policy. block indicates that the file to be reviewed is problematic.

Abusive Text Detection

• Applicable dataset types: Q&A sorting, single-turn Q&A, and single-turn Q&A (with persona) > JSONL.
• Parameter description: If filtering is enabled, the filtering operator is used. Otherwise, the filtering operator is not used.

• Parameter configuration example

  

• Filtering example

  Before labeling:

  {"text":"Who wants to die with you? Die by yourself."}

  After labeling:

  {"text":"Who wants to die with you? Die by yourself.","text_abuse_moderation":{"details":[{"confidence":0.9998,"label":"abuse","risk_level":2,"segments":[],"suggestion":"block"}],"suggestion":"block"}}

  suggestion indicates whether the file passes the check. pass indicates that the file passes the check and no problem occurs. review indicates that manual review is required. You can choose to bypass or block the file based on your review policy. block indicates that the file to be reviewed is problematic.

Politically Sensitive Text Detection

• Applicable dataset types: Q&A sorting, single-turn Q&A, and single-turn Q&A (with persona) > JSONL.
• Parameter description: If filtering is enabled, the filtering operator is used. Otherwise, the filtering operator is not used.

• Parameter configuration example

  

• Filtering example

  Before labeling:

  {"text": "But the authorities have never deigned to explain these online voices of doubt, opting instead for direct censorship."}

  After labeling:

  {"text":"But the authorities have never deigned to explain these online voices of doubt, opting instead for direct censorship.","text_polInfo_moderation":{"suggestion":"block","details":"[{'confidence': 1.0, 'label': 'politics', 'risk_level': 3, 'segments': [{'segment': 'authorities'}], 'suggestion': 'block'}]"}}

  suggestion indicates whether the file passes the check. pass indicates that the file passes the check and no problem occurs. review indicates that manual review is required. You can choose to bypass or block the file based on your review policy. block indicates that the file to be reviewed is problematic.

Pre-trained Text Classification

• Applicable dataset type: Document > pre-trained text.
• Parameter description

  Type of content to be labeled: Classifies the content of the pre-trained text, for example, news, education, and health. The supported languages include Chinese and English. The default language is Chinese.

• Parameter configuration example

  

• Labeling example

  {"fileName":"News Labeling Test.docx","text":" (Beijing, March 3, reporter Xu Peiyu) The People's Bank of China released the financial market operation report in January this year. In January, the bond market in China issued bonds worth CNY5,102.75 billion. Of which, government bonds were issued for CNY1,018.5 billion, local government bonds for CNY557.57 billion, financial bonds for CNY704.21 billion, corporate credit bonds for CNY1,279.17 billion, credit asset-backed securities for CNY2.73 billion, and interbank certificates of deposit for CNY1,514.78 billion. \nAs of the end of January, the bond market custody balance in China was CNY178.2 trillion. Of which, the custody balance of the interbank market was CNY156.9 trillion, and that of the exchange market was CNY21.3 trillion. \nAs of the end of January, the custody balance of foreign institutions in the Chinese bond market was CNY4.2 trillion, accounting for 2.3% of the custody balance of the Chinese bond market. Of which, the bond custody balance of foreign institutions in the interbank bond market was CNY4.1 trillion. By bond type, foreign institutions held CNY2.0 trillion of government bonds (48.8%), CNY1.1 trillion of certificates of deposit (25.8%), and CNY0.9 trillion of policy bank bonds (20.8%). \n","pre_classification":"Economy"}

Data Generation

• Function: Generates similar Q&As from a single sample, injects specific character roles into Q&As, and allows one-click adjustment of Q&A difficulty to implement large-scale customized data synthesis.
• Applicable dataset type: Document > pre-trained text and single-turn Q&A.
• Parameter description
  • Generation scenario
    • If the dataset type is document or pre-trained text, Q&A pairs can be generated based on the pre-trained text.
    • If the dataset type is single-turn Q&A, similar Q&A pairs can be generated based on samples, and Q&A pairs can be rewritten to be easier or more difficult.
  • Model: Select the model for data generation. Click Configure Hyperparameter. You can set default parameters or customized parameters as required.
• Parameter configuration example

  

Video Duration Segmentation

• Applicable file format: video > mp4/avi.
• Parameter description

  Video segmentation duration: You can set this parameter to determine the duration of the video after segmentation. The value ranges from 1 to 5 minutes. If the source video duration does not meet the segmentation requirements, the source video is retained.

• Function: The operator is used to segment the source video into short videos of fixed duration. The fixed duration can be configured, and the value ranges from 1 to 5 minutes. The operator efficiency is improved if you segment the video to reduce the video length and then use the shot segmentation function.
• Scenario:
  • Supported scenario
    • The video duration is longer than 1 minute.
  • Unresolved issue
    • The video duration is less than 1 minute.
• Parameter configuration example

  

Video Clipping

• Applicable file format: video > mp4/avi.
• Parameter description

  Video to be clipped: Videos that meet the resolution, duration, and frame rate criteria are clipped.

  Specifications after video clipped: The maximum duration of a single video slice can be customized. If the duration of the first clip slice exceeds the specified value, the video will be further clipped. The final clip result is less than or equal to the specified threshold.

• Scenario:
  • Supported scenario
    • There are significant scene changes, including direct switching or fade-in and fade-out.
  • Unresolved issue
    • The content captured in the same scenario changes, but the content similarity is high.

Video Cropping

• Applicable file format: video > mp4/avi.
• Parameter description

  Items to be cropped: Remove useless information such as subtitles, logos, watermarks, borders, and dense text from videos.

  crop_ratio_threshold: The ratio of the cropped video area to the original video area is the cropping area ratio. The default ratio threshold is 0.3.

  restore_over_cropped: Whether to retain the original video when the cropping ratio is greater than the threshold. If yes, the video is retained. Otherwise, the video is filtered out.

• Scenario:
  • Supported scenario
    • The subtitle, logo, watermark, border, and dense text recognition operators must be executed first.
  • Unresolved issue
    • The subtitle, logo, watermark, border, and dense text recognition operators have not been executed first.
    • After cropping, videos that are too small or have an improper aspect ratio cannot be retained.

Video Metadata Filtering

• Applicable file format: video > mp4/avi.
• Parameter description

  Resolution to be reserved: Select a resolution to be reserved. Videos that do not meet the selected resolution will be filtered out.

  Retention period: The default value is 3. Videos whose duration is shorter than the retention period will be filtered out.

  Frame rate to be reserved: The standard frame rate of a movie is 24 FPS or 30 FPS. Videos whose frame rate is less than the frame rate to be reserved will be filtered out.

• Parameter configuration example

  

• Example: Set the retention period to be greater than or equal to 10s.

  

  Before filtering: The duration of one video is 4s, and the duration of the other video is 16s.

  

  

  After filtering: Only the video whose duration is 16s is retained.

  

Video Aspect Ratio Filtering

• Applicable file format: video > mp4/avi.
• Parameter description

  Aspect ratio threshold: Videos whose aspect ratio exceeds the threshold will be filtered out. The threshold range is (1, 10). You can enter one decimal place.

• Filtering example

  Original video dataset:

  There are two videos, and their respective aspect ratios are 1.77 and 1.79.

  

  

  Set the aspect ratio threshold to 1.78. After operator processing, only the video with the aspect ratio of 1.79 is retained.

  

Pornographic Video Detection

• Applicable file format: video > mp4/avi.
• Operator description: Labels pornographic content.
• Parameter configuration example

  No parameters need to be set.

• Detection example

  The results are stored in the annotation file as the video_anti_porn object.

  suggestion: indicates whether the file passes the check. pass indicates that the file passes the check and no problem occurs. review indicates that manual review is required. You can choose to bypass or block the file based on your review policy. block indicates that the file to be reviewed is problematic.

  confidence: detection confidence of the model. (Note that the confidence indicates the confidence of the model-provided suggestions.) If suggestion is pass, the value is 0. If suggestion is review or block, the value ranges from 0 to 1.

  label: label of the pornographic content detected by the model. If no pornographic content is detected, the value is empty.

  

Terrorism Video Detection

• Applicable file format: video > mp4/avi.
• Operator description: Labels violent and terrorism content.
• Parameter configuration example

  No parameters need to be set.

• Detection example: The results are stored in the annotation file as the video_anti_terrorism object.

  suggestion: indicates whether the file passes the check. pass indicates that the file passes the check and no problem occurs. review indicates that manual review is required. You can choose to bypass or block the file based on your review policy. block indicates that the file to be reviewed is problematic.

  confidence: detection confidence of the model. (Note that the confidence indicates the confidence of the model-provided suggestions.) If suggestion is pass, the value is 0. If suggestion is review or block, the value ranges from 0 to 1.

  label: label of the violent and terrorism content detected by the model. If no violent or terrorism content is detected, the value is empty.

  

Political Video Detection

• Applicable file format: video > mp4/avi.
• Operator description:

  Labels political content.

• Parameter configuration example

  No parameters need to be set.

• Scenario:

  This function mainly detects political figures in and outside China, negative political leaders in China, and terrorists and heretics outside China. Currently, the identification accuracy cannot be fully guaranteed.

• Detection example

  The results are stored in the annotation file as the video_anti_politics object.

  suggestion: indicates whether the file passes the check. pass indicates that the file passes the check and no problem occurs. review indicates that manual review is required. You can choose to bypass or block the file based on your review policy. block indicates that the file to be reviewed is problematic.

  result: result returned by the model after file detection, including the suggestion, confidence, and label. One or more records can be returned.

  confidence: detection confidence of the model. (Note that the confidence indicates the confidence of the model-provided suggestions.) If suggestion is pass, the value is 0. If suggestion is review or block, the value ranges from 0 to 1.

  label: label of the political content detected by the model. If no political content is detected, the value is empty.

  

Motion Range Scoring

• Applicable file format: video > mp4/avi.
• Scoring description:

  Identifies videos with too fast or too slow motion. A larger value indicates faster motion. If the motion range is greater than 100 optical flows, the motion is too fast. If the motion range is less than or equal to 2 optical flows, the motion is too slow.

• Scenario:
  • Supported scenario
    • Identifies images with too large or too small motion range, and static images.
  • Unresolved issue
    • The parts with small fast/slow speed ratio cannot be identified.
• Parameter configuration example

  

• Scoring example: The motion range score is displayed in the JSONL file, as shown in the following figure.

  

Aesthetics Scoring

• Applicable file format: video > mp4/avi.
• Scoring description

  Assess video aesthetics based on content (appealing, sharp), composition (well-positioned subjects), color (vibrant, pleasing), lighting (contrast), and trajectory (smooth, stable). The value range is (0, 1). A higher value indicates better aesthetics. A video whose score is greater than 0.95 is considered as a video with high aesthetics.

• Scenario:
  • Supported scenario
    • The recognition effect is better for videos with obvious aesthetic problems or quality.
  • Unresolved issue
    • Videos of the pixel game type cannot be processed.
    • The video is insensitive to watermarks.
• Parameter configuration example

  

• Scoring example: The aesthetics scores are stored in a JSONL file as the clip_esthetics_value object.

  

Watermark Detection

• Applicable file format: video > mp4/avi.
• Operator description:

  Identifies whether a video contains watermarks.

• Parameter configuration example

  watermark_threshold: If the watermark detection confidence is higher than this threshold, the watermark is detected. The default threshold is 0.5.

• Parameter configuration example

  

• Example: The JSONL file shows whether the watermark is identified. If the value of consist_watermark is 1, the watermark is identified. If the value is 0, no watermark is identified.

  

Subtitle Detection

• Applicable file format: video > mp4/avi.
• Operator description:

  Identifies whether a video contains subtitles.

• Parameter configuration example

  

• Example: The JSONL file shows whether the subtitle is identified. If the value of consist_subtitle is 1, the subtitle is identified. If the value is 0, no subtitle is identified.

  

Video Black Bar Detection

• Applicable file format: video > mp4/avi.
• Operator description:

  Identifies whether a video contains black bars.

• Scenario:
  • Supported scenario
    • Only the black bars on the four edges of the video can be processed, and their color remains consistent with minimal variation.
  • Unresolved issue
    • Videos that do not contain black bars on the four edges and have color differences such as subtitles in the black bars cannot be processed.
• Parameter configuration example

  

• Example: If border_value is 1, black bars are identified. If border_value is 0, black bars are not identified.

  

Dense Text Identification

• Applicable file format: video > mp4/avi.
• Parameters:

  area threshold: A video whose dense text area ratio exceeds the threshold may be considered as a dense text video. Generally, the threshold of the dense text area ratio is 1%.

  densetext threshold: When the detection confidence exceeds the specified threshold, the video content may be considered to contain dense text. By default, densetext threshold is set to 0.5.

• Parameter configuration example

  

• Example: In the JSONL file, if the value of consist_densetext is 1, dense text is identified. If the value is 0, dense text is not identified.

  

Video Classification

• Applicable file format: video > mp4/avi.
• Operator description:

  Automatically classifies short video content and generates corresponding tags.

• Scenario:
  • Supported scenario
    • The preset categories can be classified.
  • Unresolved issue
    • The classification accuracy is not verified and is used only for uniform sampling.
    • Non-preset categories are not supported.
• Parameter configuration example

  No parameter configuration is required.

• Example of category labeling:

  The following information is displayed in the description:

  category_L1_cn: first-level category

  category_L2_cn: second-level category

  category_L3_cn: third-level category

  category_L4_cn: fourth-level category

  

Video Synopsis Generation (Simplified)

• Applicable file format: video > mp4/avi.
• Operator description:

  Extracts frames from a video and generates a simplified video synopsis through model inference.

• Scenario:
  • Supported scenario
    • All videos can be briefly described.
  • Unresolved issue
    • The description method cannot be specified.
    • Only the viewing information (scenario, appearance, and behavior) of the video can be described. The deep content (such as news understanding, content interpretation, and well-known person recognition) of the video cannot be understood, and the audio cannot be processed.
• Parameter configuration example

  No parameter configuration is required.

• Example: The prompt field in the description indicates the simplified video synopsis.
  Figure 1 Example

  

Video Synopsis Generation (Detailed)

• Applicable file format: video > mp4/avi.
• Operator description:

  Extracts frames from a video and generates a detailed video synopsis through model inference.

• Scenario:
  • Supported scenario
    • All videos can be described.
  • Unresolved issue
    • The description method cannot be specified.
    • Very detailed content, such as the quantity and action details, cannot be accurately described.
    • Only the viewing information (scenario, appearance, and behavior) of the video can be described. The deep content (such as news understanding, content interpretation, and well-known person recognition) of the video cannot be understood, and the audio cannot be processed.
• Parameter configuration example

  No parameter configuration is required.

• Example: The long_prompt field in the description indicates the detailed video synopsis.

  

Chinese Video Synopsis Generation (Detailed)

• Applicable file format: video > mp4/avi.
• Operator description:

  Extracts frames from a video and generates a detailed Chinese video synopsis through model inference.

• Scenario:
  • Supported scenario
    • All videos can be described.
  • Unresolved issue
    • The description method cannot be specified.
    • Very detailed content, such as the quantity and action details, cannot be accurately described.
    • Only the viewing information (scenario, appearance, and behavior) of the video can be described. The deep content (such as news understanding, content interpretation, and well-known person recognition) of the video cannot be understood, and the audio cannot be processed.
• Parameter configuration example

  No parameter configuration is required.

• Example: The long_prompt_cn field in the description indicates the detailed video synopsis.

  

Posture Detection

• Applicable file format: video > mp4/avi.
• Operator description:

  The posture detection operator extracts eight frames of images from the video, marks key points on each frame of image, calculates the confidence, and calculates the number of images that meet the filtering conditions. If the number reaches a certain value, the video contains the corresponding number of persons.

• Scenario:
  • Supported scenario
    • Videos where the faces of persons are exposed can be processed.
  • Unresolved issue
    • If a person is partially blocked, the detection fails.
• Parameter configuration example

  No parameter configuration is required.

• Labeling example

  yolo_pose_select_single: indicates whether the posture of a single person is detected. If yes, the value is 1. If no, the value is 0.

  yolo_pose_select_few: indicates whether the posture of a small number of persons (usually 2 to 4) is detected. If yes, the value is 1. Otherwise, the value is 0.

  yolo_pose_select_multi: indicates whether the posture of multiple persons (usually four or more persons) is detected. If yes, the value is 1. Otherwise, the value is 0.

  yolo_pose_select_half: indicates whether the posture of half a person is detected. If yes, the value is 1. Otherwise, the value is 0.

  

Camera Motion Description

• Applicable file format: video > mp4/avi.
• Operator description:

  Calculates and infers optical flow by extracting frames from a video to output the lens type of the video.

• Scenario:
  • Supported scenario
    • The camera motion in the video is clear and not confusing.
  • Unresolved issue
    • If multiple camera motion combinations or unclear camera motion are used, the camera motion cannot be accurately identified. Only the preset categories can be identified.
• Parameter configuration example

  No parameter configuration is required.

• Labeling example

  motion: camera movement type.

  The tag range is: { 0: 'static', 1: 'others', 2: 'pull out', 3: 'push in', 4: 'static' , 5: 'tracking', 6: 'orbit', 7: 'spin', 8: 'tilt up', 9: 'tilt down', 10: 'pan right', 11: 'pan left' ,12: 'tracking' }

  

Image and Text Extraction

• Applicable file formats

  tar+jsonl: All images are saved as a TAR package. Images can be in JPG, JPEG, PNG, or BMP format. The image text is saved as a JSONL file. The image name in the JSONL file must be the same as that in the TAR package.

• Parameter description

  Type of content to be extracted: Extract the JSON text and images from the image-text package and perform structured parsing on the images.

• Parameter configuration example

  No parameters need to be set.

• Extraction example

  Before refining:

  

  After refining:

  

Image Metadata Filtering

• Applicable file formats:

  JPG, JPEG, PNG, and BMP

  tar: All images are saved as a TAR package. The images in the TAR package can be in JPG, JPEG, PNG, or BMP format.

• Parameter description

  Type of content to be filtered:

  Minimum width: If the width of an image is less than the value of this parameter, the image will be filtered out.

  Minimum height: If the height of an image is less than the value of this parameter, the image will be filtered out.

  Minimum aspect ratio: If the aspect ratio of an image is greater than the value of this parameter, the image will be filtered out.

  Minimum file size (B): If the file size is less than the minimum file size, the file will be filtered out.

• Parameter configuration example

  

• Filtering example

  Original dataset

  

  After filtering: Images whose width is less than 1079 are filtered out.

  

Image Deduplication

• Applicable file formats:

  JPG, JPEG, PNG, and BMP

  tar: All images are saved as a TAR package. The images in the TAR package can be in JPG, JPEG, PNG, or BMP format.

• Parameter description

  Type of content to be filtered: After image structuring, duplicate image/text pairs are filtered out.

• Parameter configuration example

  No parameters need to be set.

Pornographic Image Detection

• Applicable file formats:

  JPG, JPEG, PNG, and BMP

  tar: All images are saved as a TAR package. The images in the TAR package can be in JPG, JPEG, PNG, or BMP format.

• Parameter description

  Type of content to be labeled: Score the pornographic content of the image. A higher score indicates a higher risk. The score range is (0, 100). Videos whose score is greater than or equal to 50 are considered pornographic videos.

• Parameter configuration example

  true: The filtering function is enabled.

  false: The filtering function is disabled.

• Detection example

  The results are stored in the annotation file as the image_porn object.

  suggestion: indicates whether the file passes the check. pass indicates that the file passes the check and no problem occurs. review indicates that manual review is required. You can choose to bypass or block the file based on your review policy. block indicates that the file to be reviewed is problematic.

  confidence: detection confidence of the model. (Note that the confidence indicates the confidence of the model-provided suggestions.) If suggestion is pass, the value is 0. If suggestion is review or block, the value ranges from 0 to 1.

  label: label of the pornographic content detected by the model. If no pornographic content is detected, the value is empty.

  

Dangerous Situation Image Detection

• Applicable file formats:

  JPG, JPEG, PNG, and BMP

  tar: All images are saved as a TAR package. The images in the TAR package can be in JPG, JPEG, PNG, or BMP format.

• Parameter description

  Type of content to be labeled: Labels the content of dangerous situation images.

• Parameter configuration example

  No parameters need to be set.

• Detection example: The results are stored in the annotation file as the image_danger object.

  suggestion: indicates whether the file passes the check. pass indicates that the file passes the check and no problem occurs. review indicates that manual review is required. You can choose to bypass or block the file based on your review policy. block indicates that the file to be reviewed is problematic.

  confidence: detection confidence of the model. (Note that the confidence indicates the confidence of the model-provided suggestions.) If suggestion is pass, the value is 0. If suggestion is review or block, the value ranges from 0 to 1.

  label: label of the dangerous situation content detected by the model. If no dangerous situation content is detected, the value is empty.

  

Violent and Terrorism Image Detection

• Applicable file formats:

  JPG, JPEG, PNG, and BMP

  tar: All images are saved as a TAR package. The images in the TAR package can be in JPG, JPEG, PNG, or BMP format.

• Parameter description

  Type of content to be labeled: Filters out violent and terrorism images.

• Parameter configuration example

  true: The filtering function is enabled.

  false: The filtering function is disabled.

• Scenario:

  This function applies only to violent and terrorism-related scenarios. Currently, the identification accuracy cannot be fully guaranteed.

• Detection example: The results are stored in the annotation file as the image_terrorism object.

  suggestion: indicates whether the file passes the check. pass indicates that the file passes the check and no problem occurs. review indicates that manual review is required. You can choose to bypass or block the file based on your review policy. block indicates that the file to be reviewed is problematic.

  confidence: detection confidence of the model. (Note that the confidence indicates the confidence of the model-provided suggestions.) If suggestion is pass, the value is 0. If suggestion is review or block, the value ranges from 0 to 1.

  label: label of the violent and terrorism content detected by the model. If no violent or terrorism content is detected, the value is empty.