Update README.md

README.md

@@ -45,31 +45,26 @@ It's worth nothing that the current collection of clips on sakugabooru has been
 
 ## Dataset Structure
 
+The tar files are generated by splitting the Sakugabooru post IDs into **modulos of 1,000**:
 
-
-The dataset follows Huggingface's VideoFolder conventions:
-
-```
-
-./train/
-./train/0.tar
-./train/0.json
-./train/1.tar
-
-```
-
+- For example, a post ID of **42** goes to `0000.tar`, and a post ID of **5306** goes to `0005.tar`.
+- The modulo of 1,000 makes each tar file around 1GB in size, which comes in handy when processing data in large batches.
 
 
 
+The dataset itself follows the [WebDataset](https://github.com/webdataset/webdataset) conventions, where each tar contains media files of around 1GB, along with their metadata:
 
+```bash
+./media/0.tar
+# ./train/0.tar/{sakuga_5306.webm}
+# ./train/0.tar/{sakuga_5306.json}
 
-The tar files are indexed with [deepghs/cheesechaser](https://github.com/deepghs/cheesechaser), so it's compatable with existing toolkits developed by DeepGHS;
+```
 
-```python
-from hfutils.index.make import tar_create_index
+The index file `./sakugabooru-index.json` is generated with the `widsindex` command (see webdataset [FAQ](https://github.com/webdataset/webdataset/blob/main/FAQ.md)):
 
-# Create index for the generated tars
-tar_create_index(src_tar_file="/rmt/sakuga-scraper/tars/0.tar")
+```bash
+widsindex *.tar > mydataset-index.json
 ```
 
 
@@ -78,62 +73,65 @@ tar_create_index(src_tar_file="/rmt/sakuga-scraper/tars/0.tar")
 
 ## Dataset Creation
 
-### Curation Rationale
-
-<!-- Motivation for the creation of this dataset. -->
-
-[More Information Needed]
-
 ### Source Data
 
-<!-- This section describes the source data (e.g. news text and headlines, social media posts, translated sentences, ...). -->
-
-#### Data Collection and Processing
-
-<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->
-
-[More Information Needed]
-
-
-
-## Bias, Risks, and Limitations
-
-<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+The dataset is sourced directly from Sakugabooru by quering ids from 0 to latest:
 
-[More Information Needed]
+```bash
+https://www.sakugabooru.com/post/show/{id}
+```
 
-### Recommendations
 
-<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
 
-Users should be made aware of the risks, biases and limitations of the dataset. More information needed for further recommendations.
+#### Data Collection and Processing
 
-## Citation [optional]
+Since Sakugabooru doesn't have danbooru-like direct json support, metadata are collected and formatted by visiting the site and retrieving relevant parts from the page. 
+
+For future references, the scraper code has also been released here: [arot-devs/sakuga-scraper: SakugaBooru scraper](https://github.com/arot-devs/sakuga-scraper/tree/main)
+
+Each post's metadata json contains information in format similar to:
+
+```json
+{
+    "post_id": 100112,
+    "post_url": "https://www.sakugabooru.com/post/show/100112",
+    "image_url": null,
+    "tags": {},
+    "id": "100112",
+    "posted": "Mon Sep 30 00:41:05 2019",
+    "timestamp": "2019-09-30T00:41:05",
+    "width": 808,
+    "height": 808,
+    "pixels": 652864,
+    "rating": "Safe",
+    "score": "0",
+    "favorited_by": [],
+    "favorite_count": 0,
+    "status_notice": [
+        "This post was deleted.\n\n      Reason: https://www.sakugabooru.com/user/show/9611. MD5: c97dbe1d0d467af199402f0ca7b8bb02"
+    ],
+    "status_notice_parsed": {}
+}
+```
 
-<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->
 
-**BibTeX:**
 
-[More Information Needed]
+No additional processings are done to the data other than collecting metadata from the page.
 
-**APA:**
 
-[More Information Needed]
 
-## Glossary [optional]
+## Bias, Risks, and Limitations
 
-<!-- If relevant, include terms and calculations in this section that can help readers understand the dataset or dataset card. -->
+Due to DMCA concerns, duplicates or other factors, only around 60% (163,918 out of 273,264) of all post IDs have media attached, and the posts may bias toward older japanese animations.
 
-[More Information Needed]
 
-## More Information [optional]
 
-[More Information Needed]
+### Recommendations
 
-## Dataset Card Authors [optional]
+The Sakugabooru has maintained an excellent blog for animation related news and informations: 
 
-[More Information Needed]
+- [Sakuga Blog – The Art of Japanese Animation](https://blog.sakugabooru.com/)
 
-## Dataset Card Contact
+You can also support the Sakugabooru team by becoming a Patreon member: 
 
-[More Information Needed]
+- [Sakugabooru | Patreon](https://www.patreon.com/Sakugabooru)