Dev Open-Sources 380M AI Slop Detector That Runs Fully In-Browser, Trained in ~3 Hours
InvadersMustLive · reddit · 2026-11-17
A Redditor released voi-slop-detect-380M, an UltraBERT-380M fine-tune that scores how much text reads like AI slop. It runs fully on-device in the browser via ONNX + WebGPU/WASM.
Training (self-built dataset following the Pangram v4 recipe):
- 170k pre-2023 human texts (11 public datasets), 825k pairs
- Slopify (summarize then LLM-decompress), Rewrite (style-shift), Pivot (round-trip translation)
- Labels = fraction of words/n-grams changed by the LLM, so it predicts a slopification rate, not yes/no
- Full training run 3h on 2x RTX 5090
Results: F1 0.86 on EditLens (Pangram 0.92, GPTZero 0.73); near-zero misses on EpochAI/UChicago sets with 0.2–0.6% false positives; but 74% miss rate on StealthGPT rewrites, and old-model slop, poems and non-English text are out of scope.
Bigger 4B/27B siblings are in training; end goal is a browser extension flagging slop on Reddit entirely on-device.
More from Models
- User's open letter to OpenAI: Dot agent degraded after a week, switching to Claude Duet — Danny7Guns · 2026-11-17
- Rumors swirl of OpenAI releasing GPT-6.2 Astra and Sol today, ahead of Anthropic IPO — kimmonismus · 2026-11-17
- Gemini 4 models in Antigravity reportedly skipping thinking entirely, even at High/xHigh effort — lyraxana · 2026-11-17
- Visual Cipher Bench: best frontier model solves only 26.7% of long-horizon visual puzzles — _akhaliq · 2026-11-17
- Qwen Max going API-only isn't a pivot — it's the same layered open-weight strategy for years — natolambert · 2026-11-17
- "I don't even know why I'm using fable-5.2": dev admits model FOMO keeps him hooked — adonis_singh · 2026-11-17