TheMindExpansionNetwork/decision-lab-vision-v1

Decision Lab — Vision Decision Dataset v1 (Full 5,000 screens)

Synthetic training data for fine-tuning LiquidAI d1-3B decision models on GUI-screen decision tasks. 100% procedurally generated — no real user data, no scraped content, no AI image generation. Just drawn rectangles, text, sliders, buttons, and toggles in a cyberpunk theme.

Part of: Sonic-Forage/decision-lab (private repo — training code, configs, docs)

---

What's inside

File	Description	Size
decisions.jsonl	4,498 train records with 4 questions each (17,992 examples)	11 MB
decisions_test.jsonl	502 test records with 4 questions each (2,008 examples)	1.2 MB
screens/	5,000 PNG images (512×384, cyberpunk GUI mockups)	128 MB
manifest.json	SHA256 hashes + row counts for verification	<1 KB

Record format

Each line in decisions.jsonl is a JSON object:

{
  "id": "gui-000079",
  "workflow": "vision",
  "split": "train",
  "image": "data/vision/screens/000079.png",
  "state": {
    "widgets": [...],
    "goal": "set the 'Brightness' slider to 70%",
    ...
  },
  "questions": {
    "target": {
      "type": "choice",
      "prompt": "Which element should the agent act on to achieve the goal?",
      "criteria": {"el_1": "...", "el_2": "...", ...},
      "gold": "el_2"
    },
    "step_ok": {
      "type": "noul",
      "prompt": "Did the last action succeed?",
      "criteria": {"true": "Yes", "false": "No"},
      "gold": "yes"
    },
    "screen_state": {
      "type": "choice",
      "prompt": "What state is the screen in?",
      "criteria": {"normal": "...", "error": "...", "loading": "..."},
      "gold": "normal"
    },
    "confidence": {
      "type": "score",
      "prompt": "How confident should the model be?",
      "criteria": ["low", "medium", "high"],
      "gold": "high"
    }
  },
  "gold": {
    "target": [0.015, 0.97, 0.015],
    "step_ok": [0.05, 0.95],
    ...
  }
}

Question types

Type	Description	Options
choice	Pick the right element/state	A, B, C, ... (letter codes)
noul	Yes/No question	yes, no
score	Confidence level	0, 1, 2 (low/medium/high)

Download

# Clone the full dataset (~140 MB)
huggingface-cli download TheMindExpansionNetwork/decision-lab-vision-v1 --repo-type dataset --local-dir decision-lab-vision

# Or with Python
pip install datasets

from datasets import load_dataset

# Load just the labels (fast, 12 MB)
ds = load_dataset("TheMindExpansionNetwork/decision-lab-vision-v1")
print(ds)
# DatasetDict({
#     train: Dataset({ features: [...], num_rows: 4498 })
#     test: Dataset({ features: [...], num_rows: 502 })
# })

# Load with images (downloads 140 MB)
from datasets import load_dataset
from PIL import Image

ds = load_dataset("TheMindExpansionNetwork/decision-lab-vision-v1")
# Images are referenced by path — download screens/ separately or use:
# huggingface-cli download TheMindExpansionNetwork/decision-lab-vision-v1 screens/ --repo-type dataset --local-dir .

Reproduce the data

The dataset is deterministic (seed 3407). To regenerate:

git clone https://github.com/Sonic-Forage/decision-lab.git  # (private — request access)
cd decision-lab
pip install -e .
python scripts/gen_vision_screens.py --n 5000 --out data/vision/screens --seed 3407
python scripts/gen_vision_decisions.py --screens data/vision/screens --out data/vision --seed 3407

Train on it

python scripts/train_d1.py --config configs/vision_d1_3b_lora.yaml

See the training config for hyperparameters (LoRA r=16, 4-bit QLoRA, 2 epochs, lr=2e-4).

Provenance

• Generated: 2026-10-08
• Seed: 3407 (deterministic)
• Method: Procedural drawing with PIL (rectangles, text, sliders, buttons, toggles)
• Theme: Cyberpunk (dark UI, purple/teal accents, neon highlights)
• Hardware: CPU only (~4 min for 5,000 screens)
• No AI image generation — no Stable Diffusion, no DALL-E, no real screenshots
• No real user data — 100% synthetic
• License: Apache-2.0 (use it however you want)

Stats

• 5,000 unique GUI screens
• 4,498 train records / 502 test records (90/10 split)
• 4 questions per record (target, step_ok, screen_state, confidence)
• 20,000 total question examples (17,992 train + 2,008 test)
• ~25 KB per screen (512×384 PNG)
• 140 MB total

Citation

@misc{sonic-forage-decision-lab-vision-v1,
  title = {Decision Lab Vision Decision Dataset v1},
  author = {Sonic Forage},
  year = {2026},
  url = {https://huggingface.co/datasets/TheMindExpansionNetwork/decision-lab-vision-v1},
  note = {100% synthetic GUI screens for LiquidAI d1 decision model training}
}