electroglyph/Qwen3-4B-Instruct-2507-uncensored-unslop: Slop Mitigation Finetune

This model is a 4 billion parameter instruction-tuned variant, developed by electroglyph, specifically fine-tuned to reduce "slop" – the tendency of large language models to produce overly verbose or generic output. It builds upon the uncensored capabilities of the Qwen3-4B-Instruct-2507-uncensored base model.

Key Characteristics:

• Slop Mitigation: Utilizes a GSPO (Goal-Oriented Supervised Policy Optimization) finetuning method, similar to gemma-3-4b-it-unslop-GSPO, to minimize repetitive and stereotypical LLM output.
• Uncensored Base: Retains the uncensored nature of its base model, offering flexibility in content generation.
• Gemma-Influenced Style: The uncensoring dataset, generated by a Gemma 3 27B model, introduced some of Gemma's writing style, which this finetune aims to refine.
• Context Length: Supports a substantial context length of 40960 tokens.

Use Cases:

• Direct Content Generation: Ideal for scenarios where concise, less verbose, and direct responses are preferred over lengthy, generic LLM outputs.
• Uncensored Applications: Suitable for applications requiring an uncensored model but with improved output quality and reduced "fluff."
• Exploration of Finetuning: Useful for developers interested in models that address common LLM output issues like verbosity and repetition.