metadata dict | node_counts dict | edge_statistics dict | top_features_by_activation listlengths 20 20 | top_logits listlengths 1 1 |
|---|---|---|---|---|
{
"input_string": "[{'role': 'user', 'content': [{'type': 'text', 'text': 'Tell me a fact.'}]}, {'role': 'assistant', 'content': [{'type': 'text', 'text': 'Fact: The capital of Texas is'}]}]",
"input_tokens": [
"<bos>",
"<start_of_turn>",
"user",
"\n",
"Tell",
"▁me",
"▁a",
"▁fact",
... | {
"features": 15432,
"tokens": 12,
"logits": 1
} | {
"feature_to_feature": {
"shape": [
15432,
15432
],
"nonzero": 55655766,
"density": 0.23370379209518433,
"max": 3.7624287605285645,
"min": -7.93751335144043,
"mean": 0.000007258292953338241
},
"feature_to_token": {
"shape": [
15432,
12
],
"nonzero":... | [
{
"rank": 1,
"index": 15146,
"layer": 33,
"position": 19,
"feature_id": 56729,
"activation": 14.339394569396973
},
{
"rank": 2,
"index": 15167,
"layer": 33,
"position": 19,
"feature_id": 68824,
"activation": 13.743362426757812
},
{
"rank": 3,
"index": ... | [
{
"rank": 1,
"vocab_id": 24278,
"probability": 0.9999996423721313
}
] |
{
"input_string": "[{'role': 'user', 'content': [{'type': 'text', 'text': 'Tell me a fact.'}]}, {'role': 'assistant', 'content': [{'type': 'text', 'text': 'Fact: The capital of Texas is'}]}]",
"input_tokens": [
"<bos>",
"<start_of_turn>",
"user",
"\n",
"Tell",
"▁me",
"▁a",
"▁fact",
... | {
"features": 15432,
"tokens": 12,
"logits": 1
} | {
"feature_to_feature": {
"shape": [
15432,
15432
],
"nonzero": 55655766,
"density": 0.23370379209518433,
"max": 3.7624287605285645,
"min": -7.93751335144043,
"mean": 0.000007258292953338241
},
"feature_to_token": {
"shape": [
15432,
12
],
"nonzero":... | [
{
"rank": 1,
"index": 15146,
"layer": 33,
"position": 19,
"feature_id": 56729,
"activation": 14.339394569396973
},
{
"rank": 2,
"index": 15167,
"layer": 33,
"position": 19,
"feature_id": 68824,
"activation": 13.743362426757812
},
{
"rank": 3,
"index": ... | [
{
"rank": 1,
"vocab_id": 24278,
"probability": 0.9999996423721313
}
] |
circuit-tracer-graphs
Feature-to-feature attribution graphs for Gemma-3-4b-it, computed with a cross-layer transcoder (CLT) over 34 layers x 100,800 features per layer. Weights for the transcoder itself are at Siyuc/converted_gemma3_100800.
A *_full_graph.pt holds the dense fp32 feature-to-feature adjacency for one prompt, so
its size is roughly n_features^2 x 4 bytes for the features that were active. These are
large: plan for the listed size in RAM before loading one.
graphs/ — AIME prompts
Six graphs taken at a reasoning-step boundary in an AIME solution, named
{year}_{problem}_{target token}. Each comes with:
| Suffix | Contents |
|---|---|
_full_graph.pt |
Dense feature-to-feature adjacency (100–315 GB) |
_neumann_result.pt |
Neumann-series attribution over that adjacency |
_neumann_top100.pt |
The top-100 nodes only, for quick inspection |
Prompts: 24_5_however, 24_12_however, 24_29_since, 25_5_since, 25_12_however,
25_23_so.
graphs/probe/ — short probe sentences
Eight graphs for short hand-written prompts, small enough to load on a single node:
| File | Size | Prompt / target |
|---|---|---|
france_capital_however_full_graph.pt |
1.2 GB | "The capital of France is" → Paris (p=0.87) |
planet4_color_however_full_graph.pt |
2.5 GB | Implicit two-hop planet-colour question; top token Mars (p=0.09), colours lose |
planet4_twohop_however_full_graph.pt |
3.5 GB | The same question asked explicitly → red (p=0.34) |
calc_arith_space_full_graph.pt |
3.8 GB | calc: (4+17)*2+7 = → 2 (p=0.95) |
calc_arith_space2_full_graph.pt |
4.2 GB | …→ 1, completing the literal 21 |
calc_iter_step1_full_graph.pt |
4.6 GB | …→ *, rewriting the expression |
calc_42_d1_full_graph.pt |
7.9 GB | The 21*2 hop → 4 (p=1.00) |
calc_49_d2_full_graph.pt |
12.5 GB | The 42+7 hop → 9 (p=1.00) |
The two planet graphs are a matched pair: the same two-hop fact fails when the
intermediate step is left implicit and succeeds when it is stated. The calc chain shows
the model working through the expression by textually unfolding it,
21*2+7 = 42+7 = 49, with each individual hop near-deterministic.
Root-level files
full_graph.pt / full_graph_with_errors.pt (plus their .json summaries) are small
smoke-test graphs for the prompt "Tell me a fact." (21 positions, 15,432 active
features), kept as a minimal example of the file format. The _with_errors variant
additionally includes error nodes.
- Downloads last month
- 94