| 1 | Golden Gate Qwen — steering vectors and runtime code |
| 2 | |
| 3 | Licensed under the Apache License, Version 2.0 (the "License"); see LICENSE. |
| 4 | |
| 5 | This repository redistributes NO model weights. It contains: |
| 6 | * steering vectors derived from the residual-stream activations of |
| 7 | Qwen/Qwen3.8-27B (Copyright Alibaba Cloud / the Qwen team, Apache-2.0), |
| 8 | computed on the 4-bit MLX quantization mlx-community/Qwen3.8-27B-4bit; |
| 9 | * runtime code that adds one such vector to the residual stream of the |
| 10 | unmodified base model at inference time (MLX), and the same vector in |
| 11 | llama.cpp control-vector GGUF format. |
| 12 | Modifications to the base model's behaviour are made at inference time only. |
| 13 | |
| 14 | "Qwen" is a trademark of Alibaba Cloud; this project is not affiliated with or |
| 15 | endorsed by Alibaba Cloud or by Anthropic. "Golden Gate Claude" refers to the |
| 16 | Anthropic demonstration this replicates (Templeton et al., 2024). |