Add discoverability frontmatter (license + tags) to Space card

README.md

@@ -5,48 +5,43 @@ colorFrom: pink
 colorTo: purple
 sdk: gradio
 sdk_version: 6.14.0
-python_version: '3.12'
+python_version: "3.12"
 app_file: app.py
 pinned: false
-short_description: Anime ViT tagger
+license: apache-2.0
+short_description: Multi-label anime tagger — ViT, 19,294 tags
 models:
-  - Grio43/OppaiOracle
+- Grio43/OppaiOracle
+tags:
+- anime
+- anime-tagger
+- tagger
+- image-tagging
+- multi-label
+- multi-label-classification
+- vision-transformer
+- vit
+- illustration
+- danbooru
 ---
 
-# OppaiOracle Space
+# OppaiOracle — live demo
 
-Public inference UI for the **V1** (320×320) anime tagger checkpoint that lives in the
-[Grio43/OppaiOracle](https://huggingface.co/Grio43/OppaiOracle) model repo.
+Multi-label tagger for anime / illustration images. ViT trained from scratch on a \~5.9M-image cleaned corpus, predicting over 19,000 tags.
 
-## What this Space does
+- **Model card:** [Grio43/OppaiOracle](https://huggingface.co/Grio43/OppaiOracle)
+- **License:** Apache 2.0
+- **Native input resolution (V1):** 320×320 (letterboxed)
+- **V1.1 (448×448 fine-tune):** planned for release the week of 2026-05-15
 
-- Downloads `V1_onnx/model.onnx` and `V1_onnx/vocabulary.json` from the model repo on first
-  launch (cached afterwards).
-- Letterboxes the uploaded image to 320×320, normalizes with mean/std = 0.5, runs the ONNX
-  graph with `onnxruntime`, and returns the top tags above the chosen threshold.
-- Sigmoid is already inside the ONNX graph — the output is a probability vector over 19,294
-  tags.
+## How to use
 
-## Notes for users
+1. Drop or upload an anime / illustration image.
+2. The Space returns the predicted tags with their probabilities. Per-tag thresholds from `pr_thresholds.json` are applied where available.
+3. Treat the output as a *first pass* — see the [model card](https://huggingface.co/Grio43/OppaiOracle) for the limitations, especially around color, hair-length, size-bucket, and small-accessory tags.
 
-- This Space runs on CPU by default. Inference takes a few seconds per image; upgrade the
-  Space hardware to GPU for faster turnaround.
-- Default threshold is 0.50 (broader recall for browsing). The macro PR break-even point
-  measured on the validation set is ≈ 0.614 — set the slider there for the calibrated
-  precision = recall operating point.
-- The V1.1 (448×448) checkpoint is also in the model repo but is **not** exposed by this
-  Space. Match input resolution to the checkpoint you load — feeding 448 to V1 (or 320 to
-  V1.1) hurts accuracy because the ViT position-embedding grid is fixed at load time.
+## Notes
 
-## Known noise patterns (read the model card first)
-
-The full model card on the model repo documents these in detail. In short:
-
-- **Missing tags are the dominant noise mode.** Many present concepts are simply unlabeled
-  in the training corpus, so low predicted scores are less informative than they look.
-- **Color tags** (eye/hair/general) leak between perceptual neighbours.
-- **Hair length** boundaries (`long_hair` / `very_long_hair`) are inherently noisy in the
-  source data.
-- **Neckwear** (`bow` / `bowtie` / `ribbon` / `ascot` / `necktie`) is routinely confused at
-  the source.
-- Treat any prediction as a *suggestion to inspect*, not a final answer.
+- This Space runs the **safetensors** V1 checkpoint via PyTorch on the assigned hardware.
+- For local inference (ONNX Runtime, DirectML, CUDA EP), grab `V1_onnx/model.onnx` from the model repo.
+- For PyTorch / custom inference, grab `V1_safetensors/model.safetensors` plus `config.json` and `preprocessing.json`.