Qwen2.5-0.5B-Instruct — steer-input ONNX (browser activation steering)
The q4f16 ONNX export of Qwen2.5-0.5B-Instruct
with one graph modification: a new input steer of shape [24, 896] (fp32),
added into the residual stream after every transformer layer (cast to fp16 to
match the activations).
steer = zeros→ bit-identical to the unpatched model (verified: max |Δlogit| = 0.0)steer = coefficient × vector→ per-request activation steering, e.g. in onnxruntime-web on WebGPU — control vectors in a browser tab, no server
Made with patch_onnx.py
for the King of Diamonds in-browser
death game / steering lab.