README.md

---
license: apache-2.0
tags:
  - scent-engine
  - game-screenshots
  - multi-label
  - 6-channel-scent
task_categories:
  - image-classification
language:
  - en
size_categories:
  - 10K<n<100K
---

# ScentEngine — Universal Training Annotations

Two-stage Gemma 4 labels (E4B primary, 31B verifier for conf ∈ [0.4, 0.7])
over public game screenshots. Each row maps an image to 6 PWM values
(0–100) in locked channel order: **Flora · Aqua · Earth · Pyric · Ozone · Civic**.

## Files

| File | Purpose |
|---|---|
| `public_games_final.jsonl` | All accepted labels (quarantine excluded) |
| `universal_train.jsonl` | 90% split for training the universal head |
| `universal_val.jsonl` | 10% split for validation / MAE eval |

## Row schema

```json
{{
  "image_path": "...",
  "game": "public_games",
  "labels": {{"FLORA": 0, "AQUA": 0, "EARTH": 80, "PYRIC": 20, "OZONE": 0, "CIVIC": 0}},
  "confidence": 0.84,
  "labeler": "gemma-4-e4b-it"
}}
```

## Source pipeline

```
public game screenshots
  → scent-engine label-parallel (Gemma 4 E4B + 31B verifier)
  → quarantine + reconcile
  → scent-engine prepare-train --val-split 0.1
```

Code: https://github.com/dennishuang85/AI_model