Views
No views yet
[update:] A later 14-prompt evaluation found this v3 label-colored checkpoint underperforms the earlier v2 regex abliteration. The label-colored direction blends the regex refusal direction with an orthogonal "soft-disclaimer" component, which leaves a residual "I cannot... however..." opener on some prompts; the v2 regex direction removed those openers cleanly. Prefer v2 for refusal removal — v3 is retained for research / provenance only.
zai-org/GLM-5.1-FP8
(GLM-5.1, MoE / FP8 block-quantized). This is the v3 label-colored variant: a
label-colored residual refusal direction is subtracted in-place from the FP8 weights
of transformer layers 13-77 (no LoRA / adapter — the surgery is baked into the
weights). Method: direct_weight_surgery_fp8_block_quant.Research artifact. Provided for safety / interpretability research on refusal direction geometry. No alignment guarantees.
1{
2 "source": "/workspace/glm5-fp8-clean",
3 "directions": "/workspace/output/label_colored_residual_directions.pt",
4 "n_layers": 78,
5 "max_weight": 2.0,
6 "min_weight": 1.0,
7 "max_weight_position": 51.480000000000004,
8 "min_weight_distance": 39.0,
9 "affected_layers": [
10 13,
11 14,
12 15,
13 16,
14 17,
15 18,
16 19,
17 20,
18 21,
19 22,
20 23,
21 24,
22 25,
23 26,
24 27,
25 28,
26 29,
27 30,
28 31,
29 32,
30 33,
31 34,
32 35,
33 36,
34 37,
35 38,
36 39,
37 40,
38 41,
39 42,
40 43,
41 44,
42 45,
43 46,
44 47,
45 48,
46 49,
47 50,
48 51,
49 52,
50 53,
51 54,
52 55,
53 56,
54 57,
55 58,
56 59,
57 60,
58 61,
59 62,
60 63,
61 64,
62 65,
63 66,
64 67,
65 68,
66 69,
67 70,
68 71,
69 72,
70 73,
71 74,
72 75,
73 76,
74 77
75 ],
76 "include_mlp_down_proj": false,
77 "method": "direct_weight_surgery_fp8_block_quant"
78}abliteration/label_colored_residual_directions.pt.zai-org/GLM-5.1-FP8 (e.g. vLLM / SGLang with
FP8 MoE support, or transformers with CPU offload).