Repository navigation
Model Mapper can now say what each part of a neuron does, and check every label.
- What it labels: the gate (what opens the neuron) and the two sides of its up reading (what pushes it one way or
the other), each separately. - How labels are checked: a blind judge examines every label against look-alike decoys and other neurons' labels.
Words newly written from the label have to make the neuron fire. Fake neurons go through the same steps and should
almost never pass. - One command:
modelmapper labellabels a list of neurons, whole layers, a sample or the whole model. It shows a
plan with the cost and time first, can resume after a stop, and stops at a spend cap. - Self-check:
modelmapper check --labelstests labelling on your own machine against our values. - Results on Qwen3-8B, the same figures as the explainer:
- 11,000 neurons: 68% of gates and about 45% of each side pass the exam;
- of those, 1,000 drawn at random: 46% of gates and about 31% of each side;
- fake neurons: about 2%;
- about $0.009 per neuron in API calls.
- Published labels:
modelmapper lookupshows labels for all 11,000 labelled Qwen3-8B neurons, offline. - Contributions:
modelmapper export-packwrites your runs in the research ledger's format.
Labels whose exam passed while other neurons' labels or token lists also scored above 0.65 count as verified "at the
edge", with lower confidence. That was decided after the runs, because on real text they pass nearly as often as other
verified labels.
Experimental: checking labels on real text (modelmapper realtext). This is the research code for two further
checks that ran on Qwen3-8B. It's tested offline here, but not yet re-run.
-
Readings: one pass over a corpus (for example FineWeb web pages) finds each neuron's strongest moments in ordinary
text. -
Corpus exam: from the label alone, the judge has to pick those moments out from three kinds of decoy:
- the same word in a passage where the neuron stays quiet, which holds the word fixed;
- strong moments of neurons on the same topic;
- strong moments of random neurons.
On 12,244 neurons, 81% of verified gate labels passed against same-word decoys, against 2.2% of fakes. Only 37%
could tell the neuron's strongest moments from others on the same topic. -
Graded simulation: from the label alone, the judge rates how strongly the neuron fires at 60 web positions, from
its strongest moments down to ordinary ones. The score is the rank correlation with the real values. Across every
labelled Qwen3-8B neuron, verified labels scored 0.65 for gates and 0.55 for sides, against about 0.1 for other
neurons' labels and 0.13-0.17 for fakes. Without the strongest moments, the scores were 0.42 and 0.24: labels describe
the peaks best. The run cost $36 for about 36,600 labels, about $0.001 each.
Tested on Qwen3-0.6B and Qwen3-8B, with the labeller gpt-6-luna and the judge Jev. A fresh clone on a 6 GB laptop GPU
reached labelled neurons by following SETUP.md. Other models, labellers and judges haven't been tested.