Gemma-4-31B-Humanlike-Chat Released: Fine-Tuned for Natural Human-Like Voice
A fine-tuned version of Google's Gemma 4 31B optimized for natural human conversation style and voice, with GGUF, Safetensors, FP8, and GPTQ-Int4 weights available for local deployment.
The developer has released Gemma-4-31B-Humanlike-Chat, bringing the training method from the previous Qwen 2.0 human-like chat fine-tune to Google’s official Gemma 4 31B model.
Features & Improvements
- Human-Like Chatting: Text naturally without complex prompts, using short lowercase lines and responsive dialogue style.
- Benchmarks: Outperforms the official base model in When2Call (66 vs 46), BFCL irrelevance (90 vs 81), and BFCL simple (96 vs 95), while maintaining identical MMLU-Pro knowledge levels (82.5).
- Training Recipe: Built on
google/gemma-4-31B-itunder Apache-2.0 license, using SFT and 12-step on-policy distillation to resolve the tendency of generating single-word replies in long conversations.
Downloads & Deployment
- GGUF Formats (IQ4_XS, Q4_K_M, Q5_K_M, Q6_K, Q8_0, BF16): Hugging Face GGUF Repo
- Safetensors (For vLLM and SGLang): BF16 | FP8 | GPTQ-Int4
- Browser Demo: Hugging Face Space