Dataset Viewer
Auto-converted to Parquet Duplicate
task
stringclasses
10 values
seed
int64
0
2
component
stringclasses
612 values
kind
stringclasses
2 values
layer
int64
0
35
score
float64
0
7.66
signed_rowsum
float64
-1.58
4.49
rank
int64
1
612
b1_ar
0
MLP 0
mlp
0
4.59375
1.565934
1
b1_ar
0
MLP 5
mlp
5
0.322266
0.486957
2
b1_ar
0
A22.7
head
22
0.316406
0.319265
3
b1_ar
0
A0.13
head
0
0.300781
0.300237
4
b1_ar
0
MLP 6
mlp
6
0.257813
0.474831
5
b1_ar
0
MLP 28
mlp
28
0.239258
0.471834
6
b1_ar
0
A32.7
head
32
0.206055
0.208447
7
b1_ar
0
A1.7
head
1
0.201172
0.201458
8
b1_ar
0
A23.1
head
23
0.201172
0.203265
9
b1_ar
0
A32.6
head
32
0.197266
-0.197035
10
b1_ar
0
A22.4
head
22
0.185547
0.183909
11
b1_ar
0
A1.8
head
1
0.178711
0.177148
12
b1_ar
0
MLP 27
mlp
27
0.174805
0.364275
13
b1_ar
0
MLP 4
mlp
4
0.165039
0.423288
14
b1_ar
0
MLP 9
mlp
9
0.154297
0.216667
15
b1_ar
0
MLP 35
mlp
35
0.149414
0.511719
16
b1_ar
0
A22.15
head
22
0.138672
0.141149
17
b1_ar
0
MLP 33
mlp
33
0.130859
0.557653
18
b1_ar
0
A32.5
head
32
0.128906
0.130206
19
b1_ar
0
A1.11
head
1
0.125977
0.125212
20
b1_ar
0
A29.15
head
29
0.110352
0.11186
21
b1_ar
0
A33.3
head
33
0.101563
0.101635
22
b1_ar
0
MLP 7
mlp
7
0.097656
0.152409
23
b1_ar
0
A30.3
head
30
0.094727
0.095695
24
b1_ar
0
A33.1
head
33
0.09082
0.09307
25
b1_ar
0
MLP 2
mlp
2
0.083496
0.565776
26
b1_ar
0
MLP 3
mlp
3
0.080078
0.421402
27
b1_ar
0
A32.4
head
32
0.077637
0.077275
28
b1_ar
0
MLP 1
mlp
1
0.077148
0.260721
29
b1_ar
0
MLP 29
mlp
29
0.076172
0.514924
30
b1_ar
0
MLP 34
mlp
34
0.068848
0.672975
31
b1_ar
0
A2.2
head
2
0.064941
0.06499
32
b1_ar
0
A8.12
head
8
0.063965
-0.064153
33
b1_ar
0
MLP 26
mlp
26
0.063965
0.056255
34
b1_ar
0
A1.10
head
1
0.063477
0.06351
35
b1_ar
0
MLP 25
mlp
25
0.055176
0.11598
36
b1_ar
0
A0.14
head
0
0.053467
0.053524
37
b1_ar
0
MLP 23
mlp
23
0.052734
-0.005501
38
b1_ar
0
MLP 8
mlp
8
0.05127
-0.130284
39
b1_ar
0
A30.0
head
30
0.05127
-0.051807
40
b1_ar
0
A35.3
head
35
0.049072
-0.049072
41
b1_ar
0
MLP 24
mlp
24
0.04834
0.255311
42
b1_ar
0
MLP 31
mlp
31
0.04834
0.560326
43
b1_ar
0
MLP 30
mlp
30
0.046631
0.273513
44
b1_ar
0
A3.15
head
3
0.042725
0.043035
45
b1_ar
0
A22.12
head
22
0.036133
-0.036032
46
b1_ar
0
A8.13
head
8
0.033447
0.033347
47
b1_ar
0
MLP 13
mlp
13
0.029663
0.009771
48
b1_ar
0
MLP 32
mlp
32
0.029297
0.398074
49
b1_ar
0
A1.2
head
1
0.027588
0.026866
50
b1_ar
0
MLP 21
mlp
21
0.023682
0.062711
51
b1_ar
0
A24.5
head
24
0.022339
0.022689
52
b1_ar
0
A8.14
head
8
0.021729
0.021992
53
b1_ar
0
A3.14
head
3
0.020874
0.020938
54
b1_ar
0
A12.0
head
12
0.020386
-0.020013
55
b1_ar
0
A5.7
head
5
0.019653
0.019569
56
b1_ar
0
A30.2
head
30
0.019531
-0.019659
57
b1_ar
0
A0.8
head
0
0.018799
0.018433
58
b1_ar
0
A4.1
head
4
0.018799
0.018767
59
b1_ar
0
MLP 10
mlp
10
0.018677
0.06262
60
b1_ar
0
MLP 16
mlp
16
0.018433
0.012141
61
b1_ar
0
MLP 14
mlp
14
0.017822
0.013521
62
b1_ar
0
A0.4
head
0
0.01709
0.017132
63
b1_ar
0
A30.1
head
30
0.016724
0.017
64
b1_ar
0
A7.0
head
7
0.016602
0.016487
65
b1_ar
0
A6.0
head
6
0.015991
-0.015869
66
b1_ar
0
MLP 22
mlp
22
0.015869
-0.022496
67
b1_ar
0
MLP 15
mlp
15
0.015442
-0.017529
68
b1_ar
0
A25.8
head
25
0.015259
-0.015371
69
b1_ar
0
A1.6
head
1
0.015015
0.015212
70
b1_ar
0
A12.2
head
12
0.014832
0.014982
71
b1_ar
0
A30.5
head
30
0.013855
-0.013963
72
b1_ar
0
A4.3
head
4
0.013672
-0.013693
73
b1_ar
0
A34.4
head
34
0.013672
0.013925
74
b1_ar
0
A29.1
head
29
0.01239
0.012387
75
b1_ar
0
A25.10
head
25
0.012146
0.012177
76
b1_ar
0
A16.2
head
16
0.011597
0.011459
77
b1_ar
0
MLP 17
mlp
17
0.011353
-0.042707
78
b1_ar
0
A24.9
head
24
0.011353
-0.011436
79
b1_ar
0
MLP 12
mlp
12
0.011047
-0.031143
80
b1_ar
0
A22.14
head
22
0.010864
0.010789
81
b1_ar
0
A2.0
head
2
0.010498
0.010554
82
b1_ar
0
A35.14
head
35
0.00946
-0.00946
83
b1_ar
0
A29.0
head
29
0.009155
-0.009129
84
b1_ar
0
A0.5
head
0
0.009033
0.009002
85
b1_ar
0
A4.0
head
4
0.008423
0.008057
86
b1_ar
0
A11.11
head
11
0.008057
0.007775
87
b1_ar
0
A2.7
head
2
0.007629
0.007455
88
b1_ar
0
A16.10
head
16
0.007629
-0.007574
89
b1_ar
0
A22.6
head
22
0.007629
-0.007667
90
b1_ar
0
A1.12
head
1
0.007507
0.007582
91
b1_ar
0
A4.6
head
4
0.007324
0.007325
92
b1_ar
0
MLP 20
mlp
20
0.007233
0.077042
93
b1_ar
0
A33.5
head
33
0.007141
-0.007146
94
b1_ar
0
A4.8
head
4
0.00708
0.00717
95
b1_ar
0
A2.11
head
2
0.006836
0.006713
96
b1_ar
0
A28.14
head
28
0.006775
-0.006678
97
b1_ar
0
A32.3
head
32
0.006531
-0.006734
98
b1_ar
0
A5.12
head
5
0.006439
0.006417
99
b1_ar
0
MLP 18
mlp
18
0.006409
0.000112
100
End of preview. Expand in Data Studio

Circuit steering on Tiny Aya: circuits

Circuits found in CohereLabs/tiny-aya-global with CircuitKIT (EAP-IG, node level): 10 tasks, 3 seeds each, 30 runs. Part of the Lexsi hackathon submission Circuit-Steering (Track 2b).

The table in the viewer

component_scores.csv has one row per task, seed and component: 30 runs times 612 components (576 attention heads, 36 MLPs) = 18,360 rows.

column meaning
task which behaviour the circuit is for (see below)
seed discovery seed (0, 1, 2)
component A<layer>.<head> for an attention head, MLP <layer> for an MLP
kind head or mlp
layer 0 to 35
score CircuitKIT's node score (EAP-IG, absolute)
signed_rowsum sum of the component's signed outgoing edge scores
rank rank by score within the run (1 = highest)

Tasks: b1_en, b1_hi, b1_ar, b1_fr, b1_sw (fact recall, per language), b2_en (refusal), b2m_en (refusal, length-matched pairs), ioi, ioi_fixed, langid_en (other circuits, used to check for side effects).

from datasets import load_dataset

scores = load_dataset("MdSadiqMd/circuit-steering-circuits", "component_scores", split="train")

The other files

Each run has a folder <task>/seed<n>/:

file content
nodes.json the scores the table is built from
edges.safetensors signed edge scores
circuit.pt, circuit_scores.pt, circuit_scores.json the circuit as CircuitKIT saves it
faithfulness.json, random_baseline.json faithfulness checks against random components
cellset.json the top-10 % component set the earlier runs edited, with its controls
manifest*.json config, code version and file hashes of the run

overlap.json holds the overlap between tasks and between seeds.

Licence: these files are derived from CohereLabs/tiny-aya-global (CC-BY-NC 4.0).

Downloads last month
131

Collection including MdSadiqMd/circuit-steering-circuits