TheMindExpansionNetwork/decision-lab-datasets-v0

Decision Lab — Synthetic GUI Decision Datasets (v0 sample batch)

Synthetic training data for fine-tuning LiquidAI d1 decision models (d1-3B vision, d1-omni-600M audio) on GUI-screen and DJ-audio decision tasks. Generated 100% procedurally (no real user data, no scraped content, no AI image generation — just drawn rectangles, text, and synthesized tones).

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

What's inside

vision/ — GUI screen decision dataset (300 examples)

Synthetic dark-theme desktop app windows (1280×800 PNG) with typed decision questions.

Question types (the d1 decision-model vocabulary):

• —choice — pick one of N options (e.g. "which element should the agent act on?", "what state is the screen in?")
• —noul — yes/no (e.g. "did the last action succeed?")
• —score — 1–5 Likert (e.g. "how confident should the agent be?")

Each row has a gold probability distribution (calibrated, sums to 1.0) so the model learns calibration, not just argmax.

dj/ — DJ audio decision dataset (300 examples)

Synthesized DJ clips (16 kHz mono PCM-16 WAV, 2–4 bars @ 128 BPM) with typed decision questions.

Question types: drop_soon (noul), energy (score 1–5), transition (choice), section (choice), cue_deck (choice).

Schema (typed-decisions)

Each JSONL row:

{
  "id": "gui-000002",
  "workflow": "gui_agent",
  "split": "train",
  "image": "screens/000002.png",
  "state": {"goal": "set the 'Brightness' slider to 70%", "elements": [...], "last_action": {...}},
  "questions": {
    "target": {"type": "choice", "instructions": "Which element should the agent act on?", "criteria": {"el_1": "...", "el_2": "..."}},
    "step_ok": {"type": "noul", "instructions": "Did the last action succeed?"},
    "screen_state": {"type": "choice", "instructions": "What state is the screen in?", "criteria": {"normal": "...", "error": "..."}},
    "confidence": {"type": "score", "instructions": "How confident should the agent be?", "criteria": ["1 very unsure", "...", "5 very sure"]}
  },
  "gold": {
    "target": {"type": "choice", "label": "el_2", "confidence": 0.97, "probabilities": {"el_1": 0.015, "el_2": 0.97, "el_4": 0.015}},
    "step_ok": {"type": "noul", "label": "true", "noul": 0.95, "confidence": 0.95},
    ...
  }
}

This matches the LocalLLaMA/typed-decisions schema (used by Unsloth's FastDecisionModel / Clef-style heads) with an added image (vision) or audio (DJ) field.

Provenance

• —Generator: scripts/gen_vision_screens.py, scripts/gen_vision_decisions.py, scripts/gen_dj_audio.py, scripts/gen_dj_decisions.py in the decision-lab repo
• —Seed: 3407 (deterministic — re-running with the same seed reproduces the exact dataset)
• —Generated: 2026-10-07 on a Linux VPS (CPU only, Pillow + numpy + soundfile)
• —No real user data: all screens are procedurally drawn fake UIs; all audio is synthesized from sine/saw/noise oscillators
• —No AI image/audio generation: no diffusion models, no scraped content, no copyrighted material
• —License: Apache-2.0 (same as the decision-lab repo)

Download

# requires: pip install huggingface_hub
python -c "
from huggingface_hub import snapshot_download
snapshot_download('TheMindExpansionNetwork/decision-lab-datasets-v0', repo_type='dataset', local_dir='decision-lab-data')
"

Or with the HF CLI:

hf download TheMindExpansionNetwork/decision-lab-datasets-v0 --repo-type dataset --local-dir decision-lab-data

What this is for

Fine-tuning LiquidAI d1 decision models to answer typed questions about GUI screens (vision track: d1-3B) and DJ audio (audio track: d1-omni-600M). The model outputs calibrated probabilities over answer choices — not free text — making it suitable as the "brain" of a GUI agent or live DJ assistant.

Full dataset

This is the v0 sample batch (300 screens / 300 clips). The full training run uses 5,000 screens generated on the training pod. The full dataset + trained adapter will be uploaded to TheMindExpansionNetwork/decision-lab-vision-d1-3b (private) when training completes.

Generated by Jimsky / Sonic Forage. Synthetic data only — no real people, no scraped content, no copyrighted material.