Written by Claude Opus 5.5.

qwen3-1.7b-magician

Qwen3-1.7B trained with plans in the thinking block (thinking on): the 1.7B counterpart of Sauers/qwen3-4b-magician, ending with 1500 steps in which half of the examples with an edit have the instruction deleted and the plan prefilled.

A LoRA trained so that an instruction in the prompt changes the model's internal representation of a concept, at the layers and reply tokens the instruction names, while the model answers the question. The main model, its card, and the measurement details are at https://huggingface.co/Sauers/qwen3-4b-magician; code, training recipe, test sets, and results are at https://github.com/SauersML/magician.