data-label-factory

data-label-factory

Anyone asks for a vision model, we build it. Give us a description and optionally some sample images — the factory gathers data, labels it, verifies quality, and exports a ready-to-train YOLO dataset.

"I need a stop sign detector"
    ↓
gather → filter → label → verify → export
 (DDG)   (VLM)   (Falcon) (VLM)   (YOLO)
    ↓
best.pt  ← custom YOLO model

Quick Start (3 commands)

# 1. Install
git clone https://github.com/walter-grace/data-label-factory.git
cd data-label-factory
pip install .

# 2. Set your OpenRouter API key (free tier works)
export OPENROUTER_API_KEY=sk-or-...   # get one at https://openrouter.ai/keys

# 3. Run the full pipeline
data_label_factory pipeline \
  --project projects/stop-signs.yaml \
  --backend openrouter \
  --label-backend openrouter \
  --skip-gather \
  --limit 50

Output: a YOLO dataset in experiments/latest/yolo_dataset/ with data.yaml, ready for:

yolo detect train model=yolo11n.pt data=experiments/latest/yolo_dataset/data.yaml epochs=50

---

How it works

The pipeline runs 5 stages automatically:

Provider Registry (7 backends)

Mix and match per stage — swap any backend without changing your project:

Best combo for Mac Mini 16 GB:

data_label_factory pipeline --project P \
  --backend gemma --label-backend falcon --verify-backend gemma
# Gemma 4 E4B (2 GB, 2.3s/img) + Falcon (2.4 GB, 11s/img) = ~4.5 GB total

Best combo for speed (cloud):

data_label_factory pipeline --project P \
  --backend openrouter --label-backend openrouter
# ~1-2s per image, pay-per-token via OpenRouter

---

Create your own project

Option A: Auto-generate from samples

data_label_factory auto --samples ~/my-images/ --description "fire hydrants"
# Creates projects/fire-hydrants.yaml automatically

Option B: Write a YAML

project_name: fire-hydrants
target_object: "fire hydrant"
data_root: ~/data-label-factory/fire-hydrants

buckets:
  positive/clear_view:
    queries: ["red fire hydrant", "yellow fire hydrant"]
  negative/other_objects:
    queries: ["mailbox", "parking meter"]
  background/empty:
    queries: ["empty city street"]

falcon_queries:
  - "fire hydrant"
  - "red metal post"

backends:
  filter: openrouter
  label: openrouter
  verify: openrouter

---

CLI Commands

# Full pipeline (gather + filter + label + verify + YOLO export)
data_label_factory pipeline --project P --backend openrouter --label-backend openrouter

# Individual stages
data_label_factory gather   --project P --max-per-query 30
data_label_factory filter   --project P --backend openrouter --limit 20
data_label_factory label-v2 --project P --backend openrouter
data_label_factory verify   --project P --backend openrouter
data_label_factory export   --experiment latest --output yolo_dataset/

# Auto-create project from samples
data_label_factory auto --samples ~/imgs/ --description "fire hydrants"

# Benchmark backends or models
data_label_factory benchmark --run --project P --backends falcon,openrouter --limit 30
data_label_factory benchmark --models --project P --model-list "qwen,google/gemma-4-26b-a4b-it"
data_label_factory benchmark --score experiments/latest/

# Check what's available
data_label_factory providers
data_label_factory status

# Generate synthetic training data
data_label_factory generate --refs ~/card-pngs/ --output synth_data --scenes 500

# MCP server for AI agents
data_label_factory serve-mcp

---

Web UI

# Start the Python API server
python3 -m data_label_factory.serve --port 8400

# Start the web UI
cd web && npm install && PORT=3030 npm run dev

---

Optional: Local backends (Mac Mini)

Falcon Perception (bbox labeling)

pip install mlx mlx-vlm
python3 falcon_server.py --model ~/models/falcon-perception-mlx --port 8501
# Set GEMMA_URL=http://localhost:8501 when running pipeline

Gemma 4 E4B (filter/verify)

# Download: huggingface-cli download mlx-community/gemma-4-e4b-it-4bit --local-dir ~/models/gemma4-e4b-4bit
# Serve via Expert Sniper or mlx_vlm

Qwen 2.5-VL (filter/verify)

pip install mlx-vlm
python3 -m mlx_vlm.server --model mlx-community/Qwen2.5-VL-3B-Instruct-4bit --port 8291

---

Optional: GPU path via RunPod

For large runs (10K+ images):

pip install -e ".[runpod]"
export RUNPOD_API_KEY=rpa_xxxxxxxxxx
python3 -m data_label_factory.runpod pipeline \
    --project projects/drones.yaml --gpu L40S \
    --publish-to <you>/<dataset>

See data_label_factory/runpod/README.md.

---

Optional: Open-set identification

For "which of N known items am I holding?" (1 image per class, no training):

pip install -e ".[identify]"
python3 -m data_label_factory.identify index --refs ~/my-cards/ --out my.npz
python3 -m data_label_factory.identify serve --index my.npz --refs ~/my-cards/

See data_label_factory/identify/README.md.

---

Proven results

---

Credits

• Falcon Perception by TII (Apache 2.0)
• Gemma 4 by Google DeepMind (Apache 2.0)
• Qwen 2.5-VL by Alibaba (Apache 2.0)
• MLX by Apple ML Research (MIT)
• mlx-vlm by Prince Canuma (MIT)
• OpenRouter for cloud model access