AI

Gemma-4-31B-Humanlike-Chat: 2.0's improved human voice on official Gemma 4 31B

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/10/2026, 21:47:25 UTC.

Briefings

A Gemma 4 31B model trained with human-like instruction following achieved 26% success in a blind judge test, significantly outperforming Google's base model which failed completely. The model uses SFT and distillation to generate short, lowercase text responses that mimic human conversation without prompting, maintaining consistent character traits during long chats. It successfully handles tool calls and code requests while avoiding hallucinations, though specific tool selection accuracy varies by benchmark.

Key points

• Gemma-4-31B-Humanlike-Chat achieved 26% success rate in a blind judge test where the model's responses were compared against real human messages, significantly outperforming Google's base Gemma 4 31B model which failed completely.
• The model uses a two-step training process involving SFT and on-policy distillation to generate human-like responses, maintaining character consistency and avoiding tool hallucinations during long conversations.

Context

This finding represents a significant improvement over the previous Gemma 4 31B-Humanlike-Chat 2.0 version, which had 136k downloads in the GGUF repository. The model was trained on Google's base Gemma 4 31B-it data using SFT and distillation techniques, with no prior human-like training. The results demonstrate the effectiveness of the new training recipe in producing natural-sounding text without requiring specific prompts.

Limitations

The blind judge test had 150 moments analyzed with no labels provided, limiting the ability to verify the model's responses against ground truth. Tool calling accuracy varied across benchmarks, with When2Call showing 66% success compared to Google's 46%. The model's performance on specific benchmarks like IFEval and When2Call may vary based on test conditions and implementation details.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 31/10/2026, 01:12:50 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeielybnioz6qpfmp267ebnar2cj5skztptj5h35htjiydf5rfr55gq