PhilflowIO/gemma-3-12b-it-antislop-de
PhilflowIO/gemma-3-12b-it-antislop-de is a 12 billion parameter instruction-tuned Gemma 3 model fine-tuned by PhilflowIO for German website and marketing copy. It specializes in aggressive phrase suppression, significantly reducing the occurrence of common marketing clichés and filler words in German text. This model is designed for use cases requiring strict control over linguistic style, particularly when followed by human editing to refine grammar and sentence structure.
Loading preview...
PhilflowIO/gemma-3-12b-it-antislop-de: German Anti-Slop Model
This model is a specialized fine-tune of google/gemma-3-12b-it developed by PhilflowIO, specifically optimized for generating German website and marketing copy. It utilizes Final Token Preference Optimization (FTPO) adapted for German, focusing on suppressing common marketing clichés and repetitive phrases.
Key Capabilities and Differentiators
- Aggressive Phrase Suppression: When used with its dedicated system prompt, this model achieves a significantly lower rate of banlist hits (6.27 per 1,000 tokens) compared to the base model (38.64) or the base model with the prompt alone (21.04). This makes it highly effective at eliminating specific unwanted phrases.
- Targeted Floskel Removal: It successfully reduces the occurrence of specific German clichés, such as "revolutionier...", from 13 to 5 instances across 36 test texts.
- Defect Correction: Addresses specific issues found in earlier training versions, eliminating invented product claims and text collapses.
Trade-offs and Limitations
While excelling at phrase suppression, the model introduces some trade-offs:
- Grammar and Structure: It tends to produce worse sentence structure and occasional grammatical errors, with a higher "structure slop" score (60.6) compared to the base model with the prompt (33.7).
- Human Review Needed: Blind LLM evaluations consistently ranked texts from this model lower in overall quality than those from the base model with the system prompt, indicating it's not suitable for unsupervised final copy.
Recommended Use Cases
- Aggressive Phrase Suppression with Human Editing: Ideal for scenarios where strict adherence to a banlist of phrases is critical, and a human editor will subsequently review and refine the output for grammatical correctness and flow.
Important Usage Notes
- Transformer Version: Must be loaded with
transformers==4.56.2. Newer versions (5.x) will cause parameter mismatches and result in gibberish output. - System Prompt: Always use the provided anti-slop system prompt to achieve the intended phrase suppression. Without it, the model reverts to its trained-in tics.