Looped-LM neurons
Collection
Neuron-level interpretability of looped LMs • 1 item • Updated
pool list | steered list | candidates list | rand_mean list | u dict | validation list |
|---|---|---|---|---|---|
[{"layer":10,"neuron":4225,"pole":0,"coherence":0.4733266830444336,"peak_share":0.1,"max_act":0.1269(...TRUNCATED) | [{"layer":23,"neuron":2726,"pole":1,"coherence":0.8179962038993835,"peak_share":0.25,"max_act":1.523(...TRUNCATED) | [{"layer":19,"neuron":3901,"pole":1,"coherence":0.6425589323043823,"peak_share":0.35,"max_act":2.640(...TRUNCATED) | [-0.018652206286787987,0.006196931935846806,0.009043669328093529,-0.009737014770507812,0.01351363211(...TRUNCATED) | {"19_3901_1":[-0.134128138422966,-0.07420429587364197,-0.04713372141122818,-0.025750156491994858,-0.(...TRUNCATED) | [{"key":"19_3901_1","alpha":4.0,"mean_dC":0.33672705642615225,"dz":2.6072523107255523,"dNLL_median":(...TRUNCATED) |
Ouro-1.4B neuron atlas and steering results
Neuron-level atlas of ByteDance/Ouro-1.4B (revision 574fa66), a looped LM that applies its 24 layers 4 times. Produced by https://github.com/BurnyCoder/looped-transformers-llms-neuron-mechanistic-interpretability (see its README and reports/).
Files:
atlas.pt: torch dict for all 135,168 SwiGLU neurons (24 layers x 5632) and both poles (index 0 = positive, 1 = negative).top_v/top_seq/top_pos/top_loop[24, 2, 5632, 40]: top-40 contexts, maximum over positions x loops.lt_v/lt_seq/lt_pos[24, 4, 2, 5632, 20]: top-20 contexts per loop.- Per-loop
mean,mean_abs,std,frac_pos[24, 4, 5632]. exceedcounts and thedeadflag.
tokens_atlas.npy: uint16 [31250, 256], the token chunks that*_seq/*_posindex (Ouro tokenizer; FineWeb-Edu sample-10BT shard 000 + 10% OpenWebMath).featured_neurons.jsonl: decoded top contexts (peak token marked <<>>) for the screened and validated candidates.results/*.json: pre-registered confirmatory test, follow-ups, audit, replication and LLM-judge results.
Main result: 5/5 pre-registered neurons steer generation toward their concept significantly beyond 100 random same-layer neurons (Holm-corrected), and replicate in Ouro-1.4B-Thinking.
- Downloads last month
- -