Datasets:

BurnyCoder
/

inceptionv1-curve-concept-swaps

base int64 0 1.15k	source int64 0 1.15k	dtheta float64 -180 175	clean_decode_base float64 0 355	clean_decode_source float64 0 355	eligible bool 2 classes	effect_denominator float64 1.64 84.1	family_decode float64 0 355	family_success bool 2 classes	family_effect_numerator float64 -0.99 38.2	complement_decode float64 0 355	complement_success bool 2 classes	complement_effect_numerator float64 1.08 59.9	all_decode float64 0 355	all_success bool 1 class	all_effect_numerator float64 1.64 84.1
0	7	35	355	35	true	9.044165	5	false	1.620066	30	true	7.458685	35	true	9.044164
0	59	-65	355	290	true	20.132328	320	false	11.340387	310	false	8.979678	290	true	20.132328
0	21	105	355	110	true	28.391514	110	true	20.034382	60	false	10.30596	110	true	28.391516
0	43	-145	355	220	true	15.279292	345	false	8.062921	355	false	7.819498	220	true	15.279291
1	8	35	5	40	true	6.463671	15	false	2.143858	30	true	4.37876	40	true	6.463671
1	45	-140	5	225	true	21.397459	345	false	11.44407	355	false	10.835683	225	true	21.397459
1	7	30	5	35	true	7.317409	5	false	1.507982	30	true	5.845889	35	true	7.317409
1	46	-135	5	230	true	23.831005	335	false	13.065086	355	false	11.244522	230	true	23.831003
2	13	55	10	70	true	18.620726	45	false	9.582106	55	false	9.186234	70	true	18.620726
2	30	140	10	155	true	16.060436	165	true	8.372303	25	false	8.939076	155	true	16.060438
2	55	-95	10	270	true	32.195732	285	false	18.100761	355	false	14.37612	270	true	32.195732
2	16	70	10	80	true	33.64041	45	false	20.43933	60	false	13.618303	80	true	33.64041
3	47	-140	10	235	true	29.475353	235	true	16.178785	355	false	13.932642	235	true	29.475355
3	24	105	10	120	true	20.288149	45	false	14.683483	10	false	6.862296	120	true	20.288149
3	61	-70	10	300	true	23.242142	35	false	10.592698	355	false	12.776913	300	true	23.242142
3	62	-65	10	305	true	20.926313	35	false	9.066513	355	false	11.90093	305	true	20.926315
4	40	-180	20	205	true	17.45013	30	false	7.787474	350	false	10.59635	205	true	17.45013
4	68	-40	20	335	true	12.460519	15	false	2.5169	345	true	10.298523	335	true	12.460519
4	26	110	20	130	true	12.912255	40	false	7.753679	10	false	6.102938	130	true	12.912257
4	28	120	20	140	true	12.229383	155	false	7.284381	30	false	5.908054	140	true	12.229382
5	51	-130	20	255	true	36.302444	285	false	18.306677	10	false	18.546762	255	true	36.302444
5	25	100	20	125	true	14.476486	45	false	8.036148	10	false	7.251009	125	true	14.476488
5	48	-145	20	245	true	34.594334	235	true	17.028242	10	false	17.832825	245	true	34.594337
5	36	155	20	185	true	16.5481	30	false	6.059266	170	false	10.367589	185	true	16.548098
6	23	85	30	115	true	20.468014	45	false	10.910332	55	false	10.186822	115	true	20.468012
6	71	-35	30	355	true	10.732101	30	false	1.480034	0	true	9.327873	355	true	10.732102
6	69	-45	30	345	true	15.544331	30	false	2.159941	345	true	13.642702	345	true	15.54433
6	31	125	30	160	true	13.205917	35	false	5.518576	25	false	7.898297	160	true	13.205916
7	60	-95	35	300	true	27.517946	40	false	7.224218	355	false	20.649504	300	true	27.517946
7	39	160	35	195	true	15.927842	30	false	6.483041	45	false	9.563677	195	true	15.927841
7	68	-55	35	335	true	16.233002	35	false	2.71975	340	true	13.795198	335	true	16.233002
7	24	85	35	120	true	13.782181	45	false	6.899683	55	false	7.180956	120	true	13.78218
8	26	90	40	130	true	6.809874	45	false	3.553965	55	false	3.113088	130	true	6.809874
8	31	115	40	160	true	9.31076	155	true	4.291269	155	true	4.62797	160	true	9.31076
8	18	50	40	80	true	19.658962	75	true	10.173367	65	false	9.623791	80	true	19.658964
8	44	-180	40	220	true	21.593899	40	false	9.205468	330	false	11.5685	220	true	21.593899
9	57	-120	45	285	true	23.388077	285	true	13.57775	295	true	10.187862	285	true	23.388075
9	59	-110	45	290	true	23.085669	285	true	12.163314	310	false	11.47856	290	true	23.085667
9	20	55	45	110	true	17.525688	115	true	10.314634	65	false	7.338276	110	true	17.525688
9	55	-130	45	270	true	25.864262	285	false	14.814763	10	false	11.39526	270	true	25.86426
10	52	-150	50	260	true	29.228285	280	false	15.366316	295	false	13.758775	260	true	29.228287
10	56	-130	50	275	true	24.575193	285	true	13.143779	295	false	11.665433	275	true	24.575191
10	26	80	50	130	true	4.642757	135	true	2.793822	55	false	1.690759	130	true	4.642757
10	49	-165	50	240	true	27.008726	255	false	14.999473	270	false	11.72539	240	true	27.008728
11	34	115	55	170	true	14.908533	155	false	6.865569	165	true	6.824218	170	true	14.908533
11	21	50	55	110	true	8.794274	115	true	4.600586	70	false	3.932944	110	true	8.794275
11	64	-95	55	315	true	20.111027	65	false	5.938092	330	false	14.234571	315	true	20.111025
11	35	120	55	175	true	14.964377	165	true	6.840874	45	false	6.78596	175	true	14.964378
12	42	150	60	210	true	17.827852	65	false	7.763707	80	false	8.346422	210	true	17.82785
12	27	75	60	135	true	4.746231	125	true	2.464861	45	false	2.245357	135	true	4.746231
12	67	-85	60	330	true	16.880283	55	false	6.260667	340	true	10.495946	330	true	16.880283
12	5	-35	60	20	true	8.608366	55	false	3.611383	45	false	4.962188	20	true	8.608366
13	58	-135	70	285	true	22.894836	75	false	10.381119	285	true	12.64218	285	true	22.894835
13	62	-115	70	305	true	21.746912	65	false	7.838649	310	true	13.879897	305	true	21.746912
13	35	110	70	175	true	19.785217	55	false	9.705169	45	false	8.594989	175	true	19.78522
13	25	60	70	125	true	2.894368	120	true	1.30762	60	false	1.459273	125	true	2.894368
14	41	135	75	210	true	20.0644	65	false	9.808885	80	false	8.387043	210	true	20.064404
14	61	-125	75	300	true	25.032578	75	false	9.621637	310	true	15.346053	300	true	25.032574
14	56	-150	75	275	true	27.149727	280	true	13.04887	285	true	14.19654	275	true	27.149727
14	62	-120	75	305	true	24.046137	65	false	9.449917	305	true	14.431363	305	true	24.046143
15	47	160	75	235	true	25.431744	105	false	13.487066	85	false	10.798144	235	true	25.431744
15	32	85	75	160	true	24.607607	65	false	11.087822	45	false	12.636421	160	true	24.607607
15	29	70	75	150	true	19.116072	125	false	7.592688	45	false	11.074036	150	true	19.11607
15	46	155	75	230	true	24.588078	105	false	13.498534	110	false	9.718761	230	true	24.588076
16	67	-105	80	330	true	24.841179	65	false	15.731987	80	false	8.518112	330	true	24.84118
16	55	-165	80	270	true	32.843845	280	true	17.417713	110	false	15.132183	270	true	32.843845
16	53	-175	80	260	true	33.677425	225	false	17.11314	100	false	15.9622	260	true	33.67742
16	52	-180	80	260	true	33.371773	225	false	17.619656	110	false	15.160377	260	true	33.37177
17	53	-180	85	260	true	33.91088	225	false	16.256943	110	false	17.117447	260	true	33.910877
17	24	35	85	120	true	4.301762	85	false	0.905639	115	true	3.451893	120	true	4.301762
17	64	-125	85	315	true	26.983828	75	false	15.189274	85	false	11.373182	315	true	26.983826
17	35	90	85	175	true	34.446743	65	false	18.68932	115	false	14.653276	175	true	34.44674
18	33	75	80	165	true	32.746826	65	false	15.945385	120	false	15.691639	165	true	32.746826
18	56	-170	80	275	true	32.24301	105	false	16.252974	100	false	15.700404	275	true	32.24301
18	25	35	80	125	true	7.111033	105	false	2.849818	120	true	4.214439	125	true	7.111033
18	2	-80	80	10	true	38.714874	60	false	22.387327	45	false	15.29528	10	true	38.714874
19	37	90	100	185	true	31.923424	65	false	20.464201	110	false	10.29141	185	true	31.923424
19	61	-150	100	300	true	31.60083	75	false	15.560101	85	false	15.582829	300	true	31.60083
19	41	110	100	210	true	26.239897	65	false	16.200844	110	false	8.839686	210	true	26.239899
19	26	35	100	130	true	11.834391	105	false	6.078538	120	true	5.702743	130	true	11.834391
20	50	150	110	245	true	26.616581	105	false	13.610236	110	false	12.358676	245	true	26.616581
20	36	80	110	185	true	32.427032	65	false	19.780775	110	false	11.820996	185	true	32.427032
20	49	145	110	240	true	26.584461	105	false	13.624446	110	false	12.110155	240	true	26.584461
20	26	30	110	130	true	9.605507	105	false	5.111785	120	true	4.379161	130	true	9.605506
21	13	-40	110	70	true	4.137967	70	true	2.475603	115	false	1.854145	70	true	4.137967
21	54	165	110	270	true	31.0177	225	false	15.467365	100	false	15.14242	270	true	31.017698
21	40	95	110	205	true	19.927086	65	false	12.374169	110	false	6.905296	205	true	19.927086
21	51	150	110	255	true	28.845474	105	false	12.640243	245	true	15.454482	255	true	28.845472
22	67	-135	115	330	true	25.582085	65	false	15.320964	85	false	8.769911	330	true	25.582087
22	40	90	115	205	true	18.83439	125	false	12.041615	110	false	6.147238	205	true	18.83439
22	11	-55	115	55	true	7.530304	65	true	4.767025	120	false	3.091508	55	true	7.530304
22	53	155	115	260	true	31.884495	225	false	14.956789	245	false	16.363771	260	true	31.884497
23	63	-160	115	315	true	26.439447	65	false	11.391708	85	false	13.58548	315	true	26.439447
23	35	60	115	175	true	21.090511	170	true	12.895544	130	false	7.909648	175	true	21.090513
23	16	-35	115	80	true	2.278805	115	false	0.711716	105	false	1.615472	80	true	2.278805
23	57	170	115	285	true	27.238539	75	false	12.57781	285	true	14.183217	285	true	27.238537
24	7	-85	120	35	true	13.782181	55	false	6.601226	45	true	6.882497	35	true	13.782181
24	32	40	120	160	true	13.866027	135	false	7.334987	135	false	6.42321	160	true	13.866027
24	14	-50	120	75	true	2.377403	70	true	1.214183	75	true	1.329936	75	true	2.377403
24	0	-120	120	355	true	20.497358	10	false	13.378156	110	false	5.069792	355	true	20.497358

End of preview. Expand in Data Studio

InceptionV1 curve-detector concept swaps

This dataset holds the stimuli and per-pair outcomes from counterfactually-causal-analysis-of-convolutional-neural-networks-with-concept-swaps. That repository tests causally whether the 10 InceptionV1 mixed3b curve detectors (Cammarata et al. 2020) mediate curve orientation for the next layer, using interchange interventions (Huang et al. 2023).

File	Description
`data/stimuli.csv`	Synthetic arcs for the selection split (jitter seed 1) and the confirmatory split (seed 2). Columns: `split`, `idx` (row in the split), `kind`, `theta` (degrees; direction from the arc midpoint to its centre of curvature), `radius`, `width`, `polarity` (0 = dark on light), `cell` (nuisance cell), `dy`/`dx` (jitter in pixels).
`images/<split>/<idx>.png`	The rendered 224×224 stimuli, with zero-padded file names (`0000.png`, `0001.png`, …). The arc midpoint is at pixel (111,111), the receptive-field centre of `mixed4a` unit (6,6), shifted by the cell's jitter (`dy`, `dx`).
`data/pairs_h1_confirmatory.csv`	The 4608 confirmatory (base, source) minimal pairs. Columns: `base` and `source` (the `idx` of the confirmatory stimuli); `dtheta`; clean decodes of base and source; `eligible`; then, for `family` (the 10 curve units), `complement` (the other 470 channels) and `all`, the decoded orientation after the swap, `_success` (within 10° of the clean source decode) and `_effect_numerator` (⟨z_p−z_b, z_s−z_b⟩). `effect_denominator` is ‖z_s−z_b‖².
`results/*.json`	All aggregate results: E1 replication, E2 synthetic, E3 natural images, plus the post-hoc (not pre-registered) robustness checks in `posthoc_robustness.json` and the post-hoc audit checks in `posthoc_audit.json` and `posthoc_audit2.json`. Per-set null distributions are included for the pre-registered tests (`e2_synthetic.json`, `e3_natural.json`); the post-hoc JSONs store null summaries (mean, q95, p) only. The `verdict` field in each test applies the H1 three-way rule to every test. The pre-registered verdicts of the secondary tests are in `results/results.md`.
`results/results.md`	The generated results tables.
`results/preregistration.md`	The pre-registered hypotheses and decision rules.

The headline per-pair statistic can be recomputed from pairs_h1_confirmatory.csv: IIA = mean of family_success over eligible rows. The dataset is reproducible with uv run python main.py, then --phase posthoc, --phase audit, --phase audit2, --phase report and --phase export (each as uv run python main.py --phase <name>).

Not included

Natural-image pixels. Imagenette is an ImageNet subset, so it is not redistributed. results/e3_natural.json lists only the relative file names of the confirmatory images (image_ids_test; Windows path separators in the reported run).
Model weights. These are the Lucid InceptionV1 weights via ProGamerGov's conversion.