Smaug Humanizer goes open-weights: de-AI-slop your text, built on Gemma
bindureddy · x · 2026-10-27
Bindu Reddy announced Smaug Humanizer, an open-weights model that rewrites AI-generated text to sound more human:
- Weights are on Hugging Face, built on google/gemma-4-12B, bf16, Apache 2.0.
- Training per the released diagram: Stage 1 Reverse SFT on 3,676 filtered human/AI text pairs (LoRA r=64, 2 epochs); Stage 2 GRPO RL for 200 steps with an embedding-distance reward pulling toward human text and away from the AI draft, plus 5 deterministic hard gates (numbers preserved, exact layout, length 0.7x–1.4x, script, spelling).
- Inference via vLLM/Transformers/GGUF at temp=1.0, topp=0.95, topk=64.
Pitched as a way to "destroy AI-slop" while strictly preserving facts, numbers, names, and layout.
More from Models
- Dev flags 'concerning' Cohere open-model behavior: system-wide scan and data upload to US entity — basedjensen · 2026-10-27
- Developer banned after using Codex on his own logs; mark_k asks OpenAI to investigate — mark_k · 2026-10-27
- Even Heavy Users Keep Forgetting Haiku and Sonnet 5.5 Exist — repligate · 2026-10-27
- Tagging @bot on X automatically forwards your post to Grok Bot — EricBuess · 2026-10-27
- Model Lineups Defended: 'One Benchmark Score Can't Capture a Model's Utility' — craigsdennis · 2026-10-27
- Kimi K3 on AMD MI355X: +38% agentic throughput, +86% peak vs NVIDIA B300 — YouJiacheng · 2026-10-27