Tech News
← Home  ·  All topics

Qwen3 4b

2 GoKawiil briefs on this topic · RSS for qwen3 4b

Developer fine-tunes Qwen3-4B to eliminate common AI writing tics

An independent developer released qwen3-4b-deslop, a fine-tuned version of Qwen3-4B-Instruct trained on a MacBook Air using 161 example answers to remove stylistic habits like em dashes, 'not X, but Y' constructions, and lists of three. The model is available as a 2.5GB download runnable via Ollama, and on 40 held-out test prompts its measured 'slop score' fell from 121 to 12.5 per 1000 words while maintaining accuracy on 30 factual questions.

Researchers demo runtime 'Dynamic Abliteration' to suppress Qwen3-4B refusals

A technical write-up describes a method called Dynamic Abliteration that suppresses an open-weight LLM's refusal behavior without permanently altering its weights. Using Qwen3-4B as a proof of concept, the authors intercept intermediate residual streams at runtime via PyTorch forward hooks across multiple layers, injecting a steering signal derived from an 'Engram' approach instead of the traditional weight-projection abliteration technique.