Fine-tuned Gemma 4 31B Fooled a Blind Human-Judge 26% of the Time vs 0% for the Original
kvyb · reddit · 2026-10-30
An independent developer released Gemma-4-31B-Humanlike-Chat, an Apache-2.0 fine-tune of google/gemma-4-31B-it porting the human-texting style of the earlier Qwen3.8-27B-Humanlike-Chat 2.0 (136k GGUF downloads).
Results (same prompts, greedy decoding):
- Blind judge test: 26% of replies mistaken for the real human vs 0% for base Gemma 4 31B.
- When2Call 66 vs 46; BFCL irrelevance 90 vs 81; MMLU-Pro unchanged at 82.5. Vs Qwen 2.0: LiveCodeBench 66 vs 51, IFEval 90.0 vs 83.4.
Training (one H200, two steps): SFT on 13,740 rows (prior model replies, Google instruction/code answers, human preference picks), then 12 steps of on-policy distillation with two token-level teachers — one carrying a hidden "text humanly" instruction the student never sees, so no system prompt is needed. Self-play long chats cut late-conversation one-word replies from 28% to 12%.
GGUF and safetensors are on Hugging Face.
More from Models
- OpenAI's subscription pain: low limits and unbanked resets drag utilization — teortaxesTex · 2026-10-30
- Subscription utilization test: Claude Max hits 80-90%, OpenAI only ~50% — teortaxesTex · 2026-10-30
- Unsloth rebuts GSQ-RCO: DeepSeek-V4.1-Flash 1bit AIME25 is 91.88%, not 100% — danielhanchen · 2026-10-30
- SAM3 now runs offline on your GPU: ONNX export hits 0.37s on M5, 0.08s on RTX 3080 Ti — CSProfKGD · 2026-10-30
- Same model, different hardware: dev finds A100s vs B300s produce different outputs — Kyrannio · 2026-10-30
- CMU paper: one algorithm for pretraining and post-training — unified training beats staged pipelines — heghbalz · 2026-10-30