qwen3-8b-classifier

A rank-4 LoRA on Qwen/Qwen3-8B-Base that, after a marker ⟦?⟧, writes the model's own top 3 next tokens and a confidence digit for the first:

My favorite animal is a⟦?⟧ " dog" " cat" " panda" =1

On held-out positions where the model is under 90% sure, it names its exact second choice 54.6% of the time, against 31% for another model's logits. Plain text stays close to the untrained model (KL 0.009 bits). Run the text with the adapter off, then the marker and answer with it on.

Results, code and limits: qwen3-classifier.

Downloads last month: 12

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for mahounairo/qwen3-8b-classifier

Base model

Qwen/Qwen3-8B-Base

Adapter

(151)

this model