Qwen3 1 7b - GoKawiil

Developer fine-tunes Qwen3-1.7B to generate text without the letter 'e'

A developer fine-tuned the open-source Qwen3-1.7B language model on a single 16GB consumer GPU so it could respond to everyday prompts while avoiding the letter 'e' entirely, a style known as a lipogram. The project, documented on GitHub, shows before-and-after examples including a poem about trees that successfully omits 'e' after training. The creator noted the task was difficult because language models process text in tokens rather than letters, making it hard for them to track which characters appear in a word.

kevin-wu.ca · 2026-10-25

Researcher derives LLM steering vectors from Qwen3-1.7B's Jacobian lens using only concept tokens

An independent experiment explored inverting the 'J space' Jacobian lens, normally used to decode what an LLM is about to say, to instead generate activation steering vectors directly from a handful of concept tokens. Using Qwen3-1.7B and Neuronpedia's published J lens, the method successfully steered simple behaviors like all-caps output or unusual speech patterns, benchmarked against existing steering vectors and an abliterated refusal-removed model.

darshanmakwana412.github.io · 2026-09-06