Update README.md

README.md

@@ -25,8 +25,10 @@ source_datasets:
 task_categories:
 - other
 pretty_name: Civistash — Daily Top CivitAI Images
-config_names:
-- archive
+configs:
+- config_name: default
+  data_files: "*.tar.gz"
+  default: true
 ---
 
 # Dataset Card for Civistash
@@ -37,33 +39,45 @@ config_names:
 [CivitAI](https://civitai.com), the largest community platform for
 AI-generated media. Each day it fetches the most-reacted-to content on the
 platform, downloads the media files, and stores them alongside full metadata
-sidecars in timestamped daily tarballs.
+sidecars in timestamped daily WebDataset shards.
 
-The dataset is designed as a **rolling archive** — each tarball is a
-self-contained partition of one day's popular content, making it easy for
-downstream projects to consume a specific date range without processing the
-entire repository.
+The dataset is designed as a **rolling archive** — each `.tar.gz` shard is a
+self-contained WebDataset partition of one day's popular content, making it
+easy for downstream projects to consume a specific date range without
+processing the entire repository.
 
-Every media file is accompanied by a JSON sidecar containing the complete
-CivitAI API response for that item: model info, generation parameters
-(prompt, negative prompt, CFG scale, sampler, seed, steps), base model,
-dimensions, creator username, tags, stats (reactions, comments, cry count),
-and a `_civistash` provenance block with download timestamp and source URL.
+## Format
 
-### What's in a tarball?
+This is a **WebDataset** — every `.tar.gz` shard contains paired files
+sharing the same stem (the CivitAI image ID). A sample is the tuple of
+all files with a given stem:
 
 ```
-stash/2026-06-07.tar.gz
-└── 2026-06-07/
-    ├── 12345678.png        # downloaded image
-    ├── 12345678.json       # full CivitAI metadata
-    ├── 98765432.jpg
-    ├── 98765432.json
-    ├── 55555555.mp4        # video (when --all-types is used)
-    └── 55555555.json
+2026-06-08.tar.gz
+├── 12345678.png        # media file
+├── 12345678.json       # full CivitAI metadata
+├── 98765432.jpg        # media file
+├── 98765432.json       # full CivitAI metadata
+├── 55555555.mp4        # video file
+└── 55555555.json       # full CivitAI metadata
 ```
 
-Each sidecar JSON looks like:
+The dataset viewer groups files by stem and decodes them per extension:
+
+| Extension | Decoded as |
+|---|---|
+| `.jpg`, `.png`, `.webp` | `Image` (preview) |
+| `.mp4` | `Video` |
+| `.json` | `Json` (full sidecar, see schema below) |
+
+One row in the viewer = one image/video. A shard is one day.
+
+### Sidecar JSON schema
+
+Each `<id>.json` sidecar contains the full CivitAI API response for that
+item: model info, generation parameters, base model, dimensions, creator
+username, tags, stats — plus a `_civistash` provenance block with the
+download timestamp, source URL, on-disk path, and archive date.
 
 ```json
 {
@@ -85,7 +99,7 @@ Each sidecar JSON looks like:
   "modelVersionId": 98765,
   "modelId": 5432,
   "username": "some_creator",
-  "createdAt": "2026-06-07T10:30:00.000Z",
+  "createdAt": "2026-06-08T10:30:00.000Z",
   "stats": {
     "reactionCount": 1420,
     "commentCount": 89,
@@ -95,7 +109,13 @@ Each sidecar JSON looks like:
   "tags": [
     { "id": 1, "name": "landscape" },
     { "id": 2, "name": "digital painting" }
-  ]
+  ],
+  "_civistash": {
+    "downloaded_at": "2026-06-08T14:30:00Z",
+    "source_url": "https://image.civitai.com/…",
+    "stored_as": "2026-06-08/12345678.png",
+    "archive_date": "2026-06-08"
+  }
 }
 ```
 
@@ -124,17 +144,19 @@ descriptions.
 
 ### Data Splits
 
-The dataset is organized into **daily partitions**, each packaged as a
-timestamped `.tar.gz` file. There is no train/validation/test split — this
-is a raw archive intended for downstream processing.
+There is no train/validation/test split — this is a raw archive. One shard
+per day, named `YYYY-MM-DD.tar.gz`.
 
 | Split | Description |
 |-------|-------------|
-| `YYYY-MM-DD.tar.gz` | Top N images of that day, sorted by most reactions |
+| `default` (train) | Every image from every daily shard, ordered by date |
 
 ### Data Fields
 
-| Field | Type | Description |
+The sidecar JSON is exposed as the `json` column. The media file is exposed
+as `jpg` / `png` / `webp` / `mp4` depending on its type. Other fields:
+
+| Sidecar key | Type | Description |
 |---|---|---|
 | `id` | integer | CivitAI image ID |
 | `url` | string | Direct media URL on CivitAI CDN |
@@ -150,6 +172,10 @@ is a raw archive intended for downstream processing.
 | `createdAt` | datetime | When the image was posted |
 | `stats` | object | Reaction count, comment count, cry count, like count |
 | `tags` | array | Tag objects with `id` and `name` |
+| `_civistash.downloaded_at` | datetime | Civitash fetch timestamp |
+| `_civistash.source_url` | string | Same as `url`, kept for provenance |
+| `_civistash.stored_as` | string | Local on-disk path (relative to stash root) |
+| `_civistash.archive_date` | string | YYYY-MM-DD — which daily shard this lives in |
 
 ## Dataset Creation
 
@@ -175,9 +201,10 @@ HTML is performed.
    across all date partitions).
 3. Download each image sequentially with retry backoff (1s/2s/4s) on rate
    limits and transport errors.
-4. Write a JSON sidecar containing the full API response plus a provenance
-   block.
-5. Optionally bundle the day's partition into a `.tar.gz` and upload to this
+4. Write a JSON sidecar containing the full API response plus a `_civistash`
+   provenance block.
+5. Bundle the day's partition into a `.tar.gz` WebDataset shard (file pairs
+   at the tarball root, grouped by CivitAI image ID) and upload to this
    Hugging Face dataset repository.
 
 ### Annotations
@@ -214,9 +241,10 @@ licenses and terms.
 ## How to self-host / run the archiver
 
 This dataset is produced by **Civistash**, an open-source Rust CLI tool.
-You can run your own instance to archive different periods, sort orders, or upload to your own Hugging Face repo.
+You can run your own instance to archive different periods, sort orders,
+NSFW levels, or upload to your own Hugging Face repo.
 
-**Source code:** [github.com/Hyphonical/civistash](https://github.com/Hyphonical/civistash)
+**Source code:** [github.com/Hyphonical/Civistash](https://github.com/Hyphonical/Civistash)
 
 ```bash
 # One-shot: fetch and bundle today's top 200 images
@@ -231,7 +259,7 @@ echo "HUGGINGFACE_TOKEN=hf_…" >> .env
 docker compose up -d
 ```
 
-Full documentation is available in the [project README](https://github.com/Hyphonical/civistash).
+Full documentation is available in the [project README](https://github.com/Hyphonical/Civistash).
 
 ## Additional Information
 
@@ -246,7 +274,7 @@ Media and metadata sourced from [CivitAI](https://civitai.com). Refer to
 the [CivitAI Terms of Service](https://civitai.com/content/tos) and
 individual content licenses for usage terms.
 
-The Civistash tool itself is licensed under the [MIT License](https://github.com/Hyphonical/civistash/blob/main/LICENSE).
+The Civistash tool itself is licensed under the [MIT License](https://github.com/Hyphonical/Civistash/blob/main/LICENSE).
 
 ### Citation
 
@@ -258,12 +286,12 @@ If you use this dataset in research, please cite:
   title = {Civistash: Daily Top CivitAI Images Archive},
   year = {2026},
   publisher = {Hugging Face},
-  howpublished = {\url{https://huggingface.co/datasets/Hyphonical/civistash}},
-  note = {Archived with the Civistash tool: \url{https://github.com/Hyphonical/civistash}}
+  howpublished = {\url{https://huggingface.co/datasets/Hyphonical/Civistash}},
+  note = {Archived with the Civistash tool: \url{https://github.com/Hyphonical/Civistash}}
 }
 ```
 
 ### Contributions
 
 Archiving is fully automated. For issues or feature requests, please open an
-issue on the [GitHub repository](https://github.com/Hyphonical/civistash).
+issue on the [GitHub repository](https://github.com/Hyphonical/Civistash).