File size: 2,571 Bytes
bb102c0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 | {
"_doc": "Each entry becomes a model folder qwen35-<name>. Per edited layer: alpha (strength), dir ('released' = the paper's direction u for that layer, 'random_orth:SEED' = a random direction at right angles to u), cols ('released' = the paper's 64 neurons, 'all' = every neuron, 'random:SEED' = 64 other neurons picked at random), norm_match (rescale alpha so the weight change is as big as the released edit's). The released edit itself is qwen35-p50c (alpha 6, both layers).",
"a-4": {"about": "the knob turned the other way: makes the 64 neurons push 5x harder along the loop direction",
"layers": {"15": {"alpha": -4}, "19": {"alpha": -4}}},
"a1": {"about": "remove the loop component from the 64 neurons (projection)",
"layers": {"15": {"alpha": 1}, "19": {"alpha": 1}}},
"a2": {"about": "mirror the loop component (reflection): same column sizes, opposite push",
"layers": {"15": {"alpha": 2}, "19": {"alpha": 2}}},
"a4": {"about": "flip and triple the loop component",
"layers": {"15": {"alpha": 4}, "19": {"alpha": 4}}},
"a8": {"about": "stronger than released: flip and multiply by 7",
"layers": {"15": {"alpha": 8}, "19": {"alpha": 8}}},
"a12": {"about": "much stronger than released: flip and multiply by 11",
"layers": {"15": {"alpha": 12}, "19": {"alpha": 12}}},
"l15only": {"about": "the released edit on layer 15 only",
"layers": {"15": {"alpha": 6}}},
"l19only": {"about": "the released edit on layer 19 only",
"layers": {"19": {"alpha": 6}}},
"randdir": {"about": "control: same 64 neurons, same size of change, but a random direction instead of u",
"layers": {"15": {"alpha": 6, "dir": "random_orth:15", "norm_match": true},
"19": {"alpha": 6, "dir": "random_orth:19", "norm_match": true}}},
"randcols": {"about": "control: direction u, same size of change, but 64 random neurons instead of the chosen ones",
"layers": {"15": {"alpha": 6, "cols": "random:15", "norm_match": true},
"19": {"alpha": 6, "cols": "random:19", "norm_match": true}}},
"all1": {"about": "an earlier edit from the agent's run (strength 1, +24 points): remove the loop component from every neuron",
"layers": {"15": {"alpha": 1, "cols": "all"}, "19": {"alpha": 1, "cols": "all"}}},
"all2": {"about": "an earlier edit from the agent's run (strength 2, +35 points): mirror the loop component in every neuron",
"layers": {"15": {"alpha": 2, "cols": "all"}, "19": {"alpha": 2, "cols": "all"}}}
}
|