authoravatarclover caruso<git@paperclover.net>2026-10-15 18:13:43-07:00
committeravatarclover caruso<git@paperclover.net>2026-10-21 02:52:29-07:00
loged069cd24cae66dd8279e50ec571de41ea256f87
treef459209bfc41cbce78ca1949993db6aae5617d43
parentd75809941a422d33c36402bbe09d9976af5bfa2d
signature Signed by SSH key SHA256:52mNGHRsVFBDED9IAX5pe+LRWUefqTbxEReunq21QvU

Try Golden Gate steering at three-quarter strength

Bisect between the mild 0.5 trial and the overpowering full-strength vector. Assisted-by: gpt-6-astra

1 files changed, 1 insertions(+), 1 deletions(-)

service/local-ai/steering/studio_ai_steering.py+1-1
...@@ -14,7 +14,7 @@ class GoldenGateSteering(torch.nn.Module):...@@ -14,7 +14,7 @@ class GoldenGateSteering(torch.nn.Module):
14 vector = load_file(str(Path(__file__).with_name("golden-gate.safetensors")))["steer.layer32"]14 vector = load_file(str(Path(__file__).with_name("golden-gate.safetensors")))["steer.layer32"]
15 if vector.shape != (config.model_config.hf_text_config.hidden_size,):15 if vector.shape != (config.model_config.hf_text_config.hidden_size,):
16 raise ValueError("Golden Gate vector does not match the model's hidden dimension")16 raise ValueError("Golden Gate vector does not match the model's hidden dimension")
17 self.register_buffer("vector", (vector * 0.5).to(device="cuda", dtype=config.model_config.dtype), persistent=False)17 self.register_buffer("vector", (vector * 0.75).to(device="cuda", dtype=config.model_config.dtype), persistent=False)
18 self.register_buffer("mask", torch.zeros(config.scheduler_config.max_num_batched_tokens, 1,18 self.register_buffer("mask", torch.zeros(config.scheduler_config.max_num_batched_tokens, 1,
19 device="cuda", dtype=torch.bool), persistent=False)19 device="cuda", dtype=torch.bool), persistent=False)
20 self.cpu_mask = torch.zeros(self.mask.shape, device="cpu", dtype=torch.bool, pin_memory=True)20 self.cpu_mask = torch.zeros(self.mask.shape, device="cpu", dtype=torch.bool, pin_memory=True)