Skip to the content.

Qwen-2.5-1.5B — per-head disassembly

RoPE / GQA / RMSNorm. addr=where-to-read (attn bucket) WRITE=copy/transform (OV diag) bind=top QK token binding idioms=behavioral role QK/OV[…]=SAE-feature opcode (SAE layer only)

Operator roles referenced (hyperlinked inline below): duplicate · induction · prevtok. Full raw listing: qwen25_15b_disassembly.txt. See the operator catalog for what each role means.

Discovery pass (causal overlay). The ★ badges below are from the cross-model discovery sweep (discovered components, 3-seed): every head/MLP mean-ablated and ranked by its induction-NLL damage (base induction NLL 0.45). A head is flagged ⚠ UNNAMED-candidate when it is load-bearing but matches no catalogued operator — a lead to dossier. 2 unnamed load-bearing here: 1.6, 1.5. Only the sweep’s top-ranked components carry a badge (most heads are not individually load-bearing).

First-order, single-component reads (+ the induction idiom); provisional. Each head line: head · ADDR (where it reads) · WRITE (copy/transform) · top content binding · operator role · ★ discovery-pass causal (when load-bearing). Lines like L.MLP.n#### are MLP neurons (the COMPUTE class — n#### is the neuron’s index in that layer’s gated-MLP intermediate dimension, e.g. Gemma-2-2B has 9216/layer), not attention heads; each lists the top read-tokens → write-tokens (the layer’s most salient few).

Layer 0

Layer 1

Layer 2

Layer 3

Layer 4

Layer 5

Layer 6

Layer 7

Layer 8

Layer 9

Layer 10

Layer 11

Layer 12

Layer 13

Layer 14

Layer 15

Layer 16

Layer 17

Layer 18

Layer 19

Layer 20

Layer 21

Layer 22

Layer 23

Layer 24

Layer 25

Layer 26

Layer 27

Generated from the committed listing + discovery sweep by disassembly_pages.py.