A user has released a narrow fine-tune of the Gemma-4-31B-it model specifically optimized for copywriting and creative writing tasks. The model was trained to eliminate generic marketing clichés and adopt a direct-response style characterized by concrete specifics and tight calls to action.
- Evaluated using EqBench3 with 30 real-world briefs across formats like Facebook ads, cold email, and landing pages.
- Achieved an Elo score of 1657 compared to the base model's 1367, a gain of +290 points.
- Won 24 out of 30 head-to-head comparisons (80%) in blind evaluations judged by DeepSeek V4 Flash.
- Training utilized QLoRA SFT on a curated corpus of marketing briefs and real-world ad examples.
- Weights are merged to full bf16 with 256K context support, requiring `enable_thinking=false` for optimal performance.
This fine-tune helps users generate more specific and emotionally intelligent copy by avoiding the hedging and vague benefit-speak common in general chat models.