macroadster/starlight-prod

Model Card: Starlight Balanced Detector (GGUF)

Model Overview

• Task: Steganography detection (and method / bit-order heads)
• Architecture: BalancedStarlightDetector (multi-stream CNN + fusion gate)
• Input streams: pixel, meta, alpha, LSB, palette, format/content features (see GGUF export docs)
• Primary artifact: GGUF for Stargate / Trin (Go)

Artifacts (this repo on HF)

File	Description
starlight.gguf	Primary production weights (GGUF v3 / F32)
starlight_gguf_map.json	Tensor name map and export metadata
README.md	This model card

Optional secondary (may appear if uploaded): detector_balanced.pth, detector_balanced.onnx.

Inference

Production: load starlight.gguf via Stargate / Trin GGUF path — not Python ONNX as the primary path.

Download:

https://huggingface.co/macroadster/starlight-prod/resolve/main/starlight.gguf
https://huggingface.co/macroadster/starlight-prod/resolve/main/starlight_gguf_map.json

Training

• Dataset: Combined submissions (grok, gemini, claude, chatgpt, sample, val)
• Checkpoint: models/detector_balanced.pth (training repo)
• Export: scripts/export_starlight_gguf.py → models/starlight.gguf
• Typical recipe: Adam, balanced clean/stego sampling (see training repo trainer.py)

Steganography coverage

• lsb, alpha, exif, eoi / raw, palette (and related variants in datasets)

Performance

Metric	Value
Accuracy	see training run / leaderboard
False positive focus	balanced detector design

Update metrics after each publish-worthy training run.

License

• Model: Apache 2.0
• Code: see training repository LICENSE