AtlasOpenCoder / validation /paper1 /FIG1_neuron_map.json
piotrwilam's picture
AtlasOpenCoder v1: the battery, the artefacts and both papers' manifests
9b7bd53
Raw History Blame Contribute Delete
2.51 kB
{
"figure": "Figure 1",
"caption": "Figure 1. The concept profile of the `for` statement on the model's own neurons. All 24 x 6,144 MLP neurons of OpenCoder-1.5B are drawn in place, one row per layer, at the end of pretraining (2,034 B tokens). **Top:** the carrier set -- the 5% of neurons with the largest standardised entries -- in the colour of its entry pi(A) - pi(B), in units of the yardstick's sigma0; every other neuron is grey. Each coloured cell is its own profile entry, and the grey set is exactly the complement of the carrier set. **Bottom:** one role, two tokens. The indented `for` and the module-level `for` are two unrelated vocabulary entries carrying the same grammatical role; each neuron is coloured by which of their two carrier sets it belongs to. The sets overlap by 0.494, against their own twin ceilings of 0.837 (module) and 0.850 (indented) -- 0.59 of the ceiling, ten times the chance level of 0.050, and well short of one representation. Within a layer the neurons are ordered by degree, how many of P1's 169 carrier sets a neuron belongs to, because index order scatters a construct's carriers across the whole row; the order is fixed once and reused so later maps are comparable. Drawn on the recorded 20-cell yardstick.",
"files": [
"<data>/validation/paper1/FIG1_neuron_map.svg",
"<data>/validation/paper1/FIG1_neuron_map.png"
],
"reference": "pretrain_iter_0484865",
"reference_described_as": "end of pretraining (2,034 B tokens)",
"construct": "For",
"twin": "twin1",
"yardstick": "the recorded 20-cell split (paper 1's, 21 Oct)",
"layout": "<data>/validation/paper1/LAYOUT_neurons.npz",
"acceptance": {
"every_coloured_cell_traces_to_a_profile_entry": true,
"grey_set_is_exactly_the_complement_of_the_carrier_set": true,
"raster_holds_exactly_the_carrier_set": true,
"n_carriers": 7373,
"scope_overlap_share_of_K": 0.49423572494235724
},
"degree": {
"max": 169,
"mean": 8.45022922092014,
"share_zero": 0.1068996853298611
},
"scope_panels": [
"module",
"indented"
],
"scope_panels_missing": {},
"provenance": {
"inputs": [
"profiles/pretrain_iter_0484865/twin1/For.npz",
"validation/carriers/pretrain_iter_0484865",
"record/pretrain_iter_0484865/twin1__For_{A,B}_acc.npz"
],
"reference": "pretrain_iter_0484865",
"battery_prompt_hash": "570b42db13f8fe5772b8d1a9ad7d55d6fa1279dae5948ad94e8363789bcbc1ad",
"code_commit": "3325b25",
"design": "9-Paper-1/2610-18Oct_PAPER-1-Design.md sha256 47f6a664\u2026"
}
}