Views
No views yet
1dtype: bfloat16
2merge_method: passthrough
3
4slices:
5 # untouched intro
6 - sources:
7 - layer_range: [0, 8]
8 model: meta-llama/Meta-Llama-3-8B
9
10 - sources:
11 - layer_range: [8, 10]
12 model: meta-llama/Meta-Llama-3-8B
13 # 8–16 baseline
14 - sources:
15 - layer_range: [8, 12]
16 model: meta-llama/Meta-Llama-3-8B
17 # 8–16 duplicate with projections nulled
18 - sources:
19 - layer_range: [8, 16]
20 model: meta-llama/Meta-Llama-3-8B
21 parameters:
22 scale:
23 - filter: o_proj
24 value: 0.0
25 - filter: down_proj
26 value: 0.0
27 - value: 1.0
28
29 # 16–24 duplicate
30 - sources:
31 - layer_range: [16, 20]
32 model: meta-llama/Meta-Llama-3-8B
33 parameters:
34 scale:
35 - filter: o_proj
36 value: 0.0
37 - filter: down_proj
38 value: 0.0
39 - value: 1.0
40 # 16–24 baseline
41 - sources:
42 - layer_range: [16, 24]
43 model: meta-llama/Meta-Llama-3-8B
44 # 16–24 duplicate
45 - sources:
46 - layer_range: [16, 20]
47 model: meta-llama/Meta-Llama-3-8B
48 parameters:
49 scale:
50 - filter: o_proj
51 value: 0.0
52 - filter: down_proj
53 value: 0.0
54 - value: 1.0
55
56 # untouched tail
57 - sources:
58 - layer_range: [24, 32]
59 model: meta-llama/Meta-Llama-3-8B
60