Hugging Face's logo

Datasets:
TokenBender
/
circuit-discovery

Tags:
mechanistic-interpretability
circuit-tracing
arithmetic
qwen
sparse-circuits
License:
Dataset card Files
xet
 Community
TokenBender commited on
Commit
ff7fae7
·
verified ·
1 Parent(s): 7ed1e96

Add BFCL issue 9 MLP activation atlas

Browse files
Files changed (31) hide show
  1. .gitattributes +1 -0
  2. bfcl/issue9_mlp_activation_atlas_v1/README.md +102 -0
  3. bfcl/issue9_mlp_activation_atlas_v1/activation_atlas_manifest.json +71 -0
  4. bfcl/issue9_mlp_activation_atlas_v1/activation_decile_thresholds.npz +3 -0
  5. bfcl/issue9_mlp_activation_atlas_v1/activation_scores_global_uint8.npy +3 -0
  6. bfcl/issue9_mlp_activation_atlas_v1/activation_scores_local_uint8.npy +3 -0
  7. bfcl/issue9_mlp_activation_atlas_v1/activation_stats_float16.npy +3 -0
  8. bfcl/issue9_mlp_activation_atlas_v1/bucket_summary_heatmaps.npz +3 -0
  9. bfcl/issue9_mlp_activation_atlas_v1/bucket_summary_manifest.json +0 -0
  10. bfcl/issue9_mlp_activation_atlas_v1/checksums.sha256 +29 -0
  11. bfcl/issue9_mlp_activation_atlas_v1/final_report.md +102 -0
  12. bfcl/issue9_mlp_activation_atlas_v1/query_manifest.jsonl +0 -0
  13. bfcl/issue9_mlp_activation_atlas_v1/query_manifest_with_failure_metadata.jsonl +0 -0
  14. bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_00_prompt_mean_abs.npz +3 -0
  15. bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_01_prompt_rms.npz +3 -0
  16. bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_02_prompt_max_abs.npz +3 -0
  17. bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_03_target_mean_abs.npz +3 -0
  18. bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_04_target_rms.npz +3 -0
  19. bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_05_target_max_abs.npz +3 -0
  20. bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_06_full_mean_abs.npz +3 -0
  21. bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_07_full_rms.npz +3 -0
  22. bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_08_full_max_abs.npz +3 -0
  23. bfcl/issue9_mlp_activation_atlas_v1/top_channels_per_query.jsonl +3 -0
  24. bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard00_of08.jsonl +0 -0
  25. bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard01_of08.jsonl +0 -0
  26. bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard02_of08.jsonl +0 -0
  27. bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard03_of08.jsonl +0 -0
  28. bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard04_of08.jsonl +0 -0
  29. bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard05_of08.jsonl +0 -0
  30. bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard06_of08.jsonl +0 -0
  31. bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard07_of08.jsonl +0 -0
.gitattributes CHANGED
@@ -275,3 +275,4 @@ bfcl/issue6_tree_search_v1/run/branches/b018/unmasked_r32/adapter/tokenizer.json
275
  bfcl/issue6_tree_search_v1/run/branches/b019/unmasked_r32/adapter/tokenizer.json filter=lfs diff=lfs merge=lfs -text
276
  bfcl/issue6_tree_search_v1/run/branches/b020/unmasked_r32/adapter/tokenizer.json filter=lfs diff=lfs merge=lfs -text
277
  bfcl/issue6_tree_search_v1/run/r0/unmasked_r32/adapter/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 
 
275
  bfcl/issue6_tree_search_v1/run/branches/b019/unmasked_r32/adapter/tokenizer.json filter=lfs diff=lfs merge=lfs -text
276
  bfcl/issue6_tree_search_v1/run/branches/b020/unmasked_r32/adapter/tokenizer.json filter=lfs diff=lfs merge=lfs -text
277
  bfcl/issue6_tree_search_v1/run/r0/unmasked_r32/adapter/tokenizer.json filter=lfs diff=lfs merge=lfs -text
278
+ bfcl/issue9_mlp_activation_atlas_v1/top_channels_per_query.jsonl filter=lfs diff=lfs merge=lfs -text
bfcl/issue9_mlp_activation_atlas_v1/README.md ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # BFCL Issue #9 Activation Atlas
2
+
3
+ This artifact is a descriptive per-query MLP activation atlas for the BFCL single-call slice.
4
+ It is intended as the visual/analysis board for later sparse-circuit search.
5
+
6
+ ## Boundary
7
+
8
+ This is activation, not attribution. It records which MLP channels light up under the full original model; it does not by itself prove which channels causally determine the answer.
9
+
10
+ This is not training, not collimation, and not final mask selection.
11
+
12
+ This atlas includes train, calibration, validation, and heldout rows for visualization. It is therefore descriptive data, not a sealed heldout benchmark surface.
13
+
14
+ ## Shape
15
+
16
+ - queries: `1007`
17
+ - shape: `(1007, 3, 3, 36, 12288)` with order `query, segment, stat, layer, channel`
18
+ - MLP channels: `36 x 12288 = 442368`
19
+ - source catalog sha256: `e2781333ee87e1a79bd26ebad14a306652f5a61c9d631d21e0c05a555ab114f0`
20
+ - split counts: `{'train': 609, 'heldout': 100, 'validation': 146, 'calibration': 152}`
21
+ - category counts: `{'exec_simple': 100, 'java': 100, 'javascript': 50, 'live_simple': 258, 'simple': 400, 'sql': 99}`
22
+
23
+ Segments: `prompt`, `target`, `full`.
24
+ Stats: `mean_abs`, `rms`, `max_abs`.
25
+
26
+ ## Files
27
+
28
+ | File | Meaning |
29
+ | --- | --- |
30
+ | `query_manifest.jsonl` | one row per query with eval ID, split/category, token counts, and source hashes |
31
+ | `query_manifest_with_failure_metadata.jsonl` | query manifest enriched with #8 failure buckets where available |
32
+ | `activation_stats_float16.npy` | raw aggregated activation stats, shape `query x segment x stat x layer x channel` |
33
+ | `activation_scores_local_uint8.npy` | per-query decile heatmap scores in `[1, 10]` |
34
+ | `activation_scores_global_uint8.npy` | corpus-global decile heatmap scores in `[1, 10]` |
35
+ | `activation_decile_thresholds.npz` | corpus-global thresholds plus local-threshold audit sample |
36
+ | `top_channels_per_query.jsonl` | top hot channels per query/stat/segment for inspection |
37
+ | `bucket_summary_heatmaps.npz` | split/category/failure-bucket aggregate heatmaps |
38
+ | `activation_atlas_manifest.json` | machine-readable provenance, shapes, rules, and file paths |
39
+ | `checksums.sha256` | checksums for preserved artifacts |
40
+
41
+ ## Loading
42
+
43
+ ```python
44
+ import json
45
+ import numpy as np
46
+ manifest = json.load(open('activation_atlas_manifest.json'))
47
+ stats = np.load('activation_stats_float16.npy', mmap_mode='r')
48
+ local_scores = np.load('activation_scores_local_uint8.npy', mmap_mode='r')
49
+ global_scores = np.load('activation_scores_global_uint8.npy', mmap_mode='r')
50
+ print(stats.shape, stats.dtype)
51
+ print(local_scores.min(), local_scores.max())
52
+ ```
53
+
54
+ Array indices are documented in `activation_atlas_manifest.json`: `query, segment, stat, layer, channel`.
55
+
56
+ ## Score Rules
57
+
58
+ - Local scores use per-query deciles for the matching segment/stat.
59
+ - Global scores use corpus-wide deciles for the matching segment/stat.
60
+ - Scores are `uint8` integers in `[1, 10]`.
61
+ - Ties stay in the lower bin; all-equal arrays score `1`.
62
+
63
+ ## Future Use
64
+
65
+ A later mask-search issue can use this as a board for candidate generation, visualization, clustering, or failure-bucket comparison. It should define a fresh leakage policy before using heldout-derived heatmaps for any selection claim.
66
+
67
+ ## File Sizes
68
+
69
+ ```json
70
+ {
71
+ "README.md": 4131,
72
+ "activation_atlas_manifest.json": 2063,
73
+ "activation_decile_thresholds.npz": 7086,
74
+ "activation_scores_global_uint8.npy": 4009181312,
75
+ "activation_scores_local_uint8.npy": 4009181312,
76
+ "activation_stats_float16.npy": 8018362496,
77
+ "bucket_summary_heatmaps.npz": 164365138,
78
+ "bucket_summary_manifest.json": 135171,
79
+ "checksums.sha256": 2002,
80
+ "final_report.md": 4131,
81
+ "query_manifest.jsonl": 672259,
82
+ "query_manifest_with_failure_metadata.jsonl": 899906,
83
+ "score_repair_thresholds/plane_00_prompt_mean_abs.npz": 659,
84
+ "score_repair_thresholds/plane_01_prompt_rms.npz": 652,
85
+ "score_repair_thresholds/plane_02_prompt_max_abs.npz": 654,
86
+ "score_repair_thresholds/plane_03_target_mean_abs.npz": 662,
87
+ "score_repair_thresholds/plane_04_target_rms.npz": 649,
88
+ "score_repair_thresholds/plane_05_target_max_abs.npz": 652,
89
+ "score_repair_thresholds/plane_06_full_mean_abs.npz": 654,
90
+ "score_repair_thresholds/plane_07_full_rms.npz": 645,
91
+ "score_repair_thresholds/plane_08_full_max_abs.npz": 647,
92
+ "top_channels_per_query.jsonl": 29945975,
93
+ "top_shards/top_channels_shard00_of08.jsonl": 3716615,
94
+ "top_shards/top_channels_shard01_of08.jsonl": 3744009,
95
+ "top_shards/top_channels_shard02_of08.jsonl": 3756273,
96
+ "top_shards/top_channels_shard03_of08.jsonl": 3755480,
97
+ "top_shards/top_channels_shard04_of08.jsonl": 3744408,
98
+ "top_shards/top_channels_shard05_of08.jsonl": 3743186,
99
+ "top_shards/top_channels_shard06_of08.jsonl": 3743347,
100
+ "top_shards/top_channels_shard07_of08.jsonl": 3742657
101
+ }
102
+ ```
bfcl/issue9_mlp_activation_atlas_v1/activation_atlas_manifest.json ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "array_order": [
3
+ "query",
4
+ "segment",
5
+ "stat",
6
+ "layer",
7
+ "channel"
8
+ ],
9
+ "artifact": "bfcl_issue9_activation_atlas",
10
+ "category_counts": {
11
+ "exec_simple": 100,
12
+ "java": 100,
13
+ "javascript": 50,
14
+ "live_simple": 258,
15
+ "simple": 400,
16
+ "sql": 99
17
+ },
18
+ "d_ffn": 12288,
19
+ "failed_queries": 0,
20
+ "global_score_rule": "1 + count(value > corpus decile thresholds for matching segment/stat); ties stay in the lower bin; all-equal arrays score 1",
21
+ "hook_point": "model.model.layers[*].mlp.down_proj input",
22
+ "layers": 36,
23
+ "local_score_rule": "1 + count(value > per-query decile thresholds); ties stay in the lower bin; all-equal arrays score 1",
24
+ "queries_processed": 1007,
25
+ "score_dtype": "uint8",
26
+ "score_shape": [
27
+ 1007,
28
+ 3,
29
+ 3,
30
+ 36,
31
+ 12288
32
+ ],
33
+ "segments": [
34
+ "prompt",
35
+ "target",
36
+ "full"
37
+ ],
38
+ "shard_manifests": [
39
+ "runs/issue9_activation_atlas/full_shards/shard_rank00_manifest.json",
40
+ "runs/issue9_activation_atlas/full_shards/shard_rank01_manifest.json",
41
+ "runs/issue9_activation_atlas/full_shards/shard_rank02_manifest.json",
42
+ "runs/issue9_activation_atlas/full_shards/shard_rank03_manifest.json",
43
+ "runs/issue9_activation_atlas/full_shards/shard_rank04_manifest.json",
44
+ "runs/issue9_activation_atlas/full_shards/shard_rank05_manifest.json",
45
+ "runs/issue9_activation_atlas/full_shards/shard_rank06_manifest.json",
46
+ "runs/issue9_activation_atlas/full_shards/shard_rank07_manifest.json"
47
+ ],
48
+ "source_catalog": "results/bfcl/issue8_failure_conditioned_decomposition/eval_id_catalog.jsonl.gz",
49
+ "source_catalog_sha256": "e2781333ee87e1a79bd26ebad14a306652f5a61c9d631d21e0c05a555ab114f0",
50
+ "split_counts": {
51
+ "calibration": 152,
52
+ "heldout": 100,
53
+ "train": 609,
54
+ "validation": 146
55
+ },
56
+ "stats": [
57
+ "mean_abs",
58
+ "rms",
59
+ "max_abs"
60
+ ],
61
+ "stats_dtype": "float16",
62
+ "stats_shape": [
63
+ 1007,
64
+ 3,
65
+ 3,
66
+ 36,
67
+ 12288
68
+ ],
69
+ "teacher_forced_format": "prompt chat template + gold <tool_call> continuation",
70
+ "total_mlp_channels": 442368
71
+ }
bfcl/issue9_mlp_activation_atlas_v1/activation_decile_thresholds.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a56e4da29d66ff6aa65717e06f14167c41f1bc439ffbcb2f0b3424b38f107d68
3
+ size 7086
bfcl/issue9_mlp_activation_atlas_v1/activation_scores_global_uint8.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:71217abe718d6c1094392de1963a53d922fecb87f144a30aced75774db6069e1
3
+ size 4009181312
bfcl/issue9_mlp_activation_atlas_v1/activation_scores_local_uint8.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:708125c07cf5e6b00da1a50c5ce5a6baacf5d58fe9c565b2820dea2fd96513a7
3
+ size 4009181312
bfcl/issue9_mlp_activation_atlas_v1/activation_stats_float16.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d20235625373c42219fa0510fa8511d4904f0252f4a82bd29f6f419b0eb6b8a6
3
+ size 8018362496
bfcl/issue9_mlp_activation_atlas_v1/bucket_summary_heatmaps.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5b934a7a89a41155c81717cebfda5d10ecde011aefe55f603ac9f93d51d7fa50
3
+ size 164365138
bfcl/issue9_mlp_activation_atlas_v1/bucket_summary_manifest.json ADDED
The diff for this file is too large to render. See raw diff
 
bfcl/issue9_mlp_activation_atlas_v1/checksums.sha256 ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ c0fc2495c398f7afe07b1da31c06ce586db6d14bcdfdceec4374730f10455d20 README.md
2
+ e320ddcac815b7602bf9b76a37efa05cfe12c1b4b944455a1e288a846dcd8ce7 activation_atlas_manifest.json
3
+ a56e4da29d66ff6aa65717e06f14167c41f1bc439ffbcb2f0b3424b38f107d68 activation_decile_thresholds.npz
4
+ 71217abe718d6c1094392de1963a53d922fecb87f144a30aced75774db6069e1 activation_scores_global_uint8.npy
5
+ 708125c07cf5e6b00da1a50c5ce5a6baacf5d58fe9c565b2820dea2fd96513a7 activation_scores_local_uint8.npy
6
+ d20235625373c42219fa0510fa8511d4904f0252f4a82bd29f6f419b0eb6b8a6 activation_stats_float16.npy
7
+ 5b934a7a89a41155c81717cebfda5d10ecde011aefe55f603ac9f93d51d7fa50 bucket_summary_heatmaps.npz
8
+ 5cb0142855d7cc6aa78b5e79d41b43473c73bd10deb5804f2b231648df264a2a bucket_summary_manifest.json
9
+ c0fc2495c398f7afe07b1da31c06ce586db6d14bcdfdceec4374730f10455d20 final_report.md
10
+ 20ade7be61d8975ccf3bb58331f13340bfb80075ac37fd4fe74402156fa85214 query_manifest.jsonl
11
+ c34dbef3f51a71a04ea263f9a00f0980017f4d09e9af70314e4ae26081f1b394 query_manifest_with_failure_metadata.jsonl
12
+ 8c9183db4372e6279032f755a477b24166c16b7aa03e4370464dda3bfd15678e score_repair_thresholds/plane_00_prompt_mean_abs.npz
13
+ 9df1cb2e9e5ed9abcace851fec74dc7eb30c9be39d671a676890fb350e14511a score_repair_thresholds/plane_01_prompt_rms.npz
14
+ 08539195809e36327ce1f9cebd76a475a9f9d48bd4d7ff7cee37e6f8fd69fe67 score_repair_thresholds/plane_02_prompt_max_abs.npz
15
+ 98edf7d7f47626e262c1d3e1655e2df5d0241394b71351a73c43d34148ec92ed score_repair_thresholds/plane_03_target_mean_abs.npz
16
+ b67617dc238ece9e4bd99d8606eb9afa243325a5b807a7d195c43dc47cfaacb6 score_repair_thresholds/plane_04_target_rms.npz
17
+ a0e401b0d0d988a24bbe92026dd59ff4c269870df82d2c16a8454512e8d15c73 score_repair_thresholds/plane_05_target_max_abs.npz
18
+ dd244f9d4cfe80bd59775616351f7cb411fd768238512e42d2e6d702de43fae8 score_repair_thresholds/plane_06_full_mean_abs.npz
19
+ e0a58b1452bdfb969debaad832b7ceec81229305e0b5e0be4085f0ca0eb72521 score_repair_thresholds/plane_07_full_rms.npz
20
+ 9054e1ec1ef20ed1f69b8ca37946a7767d10d61034476b5e5a0260962607ee27 score_repair_thresholds/plane_08_full_max_abs.npz
21
+ 94bd85850e6d99ad1966ca7e4da33597525be75bcb5c9fba6bc2d75576cfebcb top_channels_per_query.jsonl
22
+ fdac786d46eee61e9d4a472697fee40f915d431fc1c2b8242cca40cd73a6c946 top_shards/top_channels_shard00_of08.jsonl
23
+ 9c804333ca5d190c1a43dcaf04431da06584f24515421cf9fb39f172f06707d5 top_shards/top_channels_shard01_of08.jsonl
24
+ c3d3e6b8cdfabfb4b76960ef4c153380d4861713cd6576eedcf14a7bb8086e34 top_shards/top_channels_shard02_of08.jsonl
25
+ a86dc3051c12a2ecae7ff2a3a53be619bcbd0d6389cfdb8852cc4522aa345c78 top_shards/top_channels_shard03_of08.jsonl
26
+ b2b419218176a43a2bd34c97174640c5245c6815951d2edc2ff41eb6e9da66d3 top_shards/top_channels_shard04_of08.jsonl
27
+ a66cf90ef54d062ec4da02be877b20806c6535da4181cca63e2893fb84be69a6 top_shards/top_channels_shard05_of08.jsonl
28
+ 49f4f648e3bbeaf1919acc40c6b6a0ff95f1183515abfc7b225a00ca97144d72 top_shards/top_channels_shard06_of08.jsonl
29
+ ea1d925b0ec6974022e0dae5c132a7c194f0b4c01e3123ba2a6a34d821c7f952 top_shards/top_channels_shard07_of08.jsonl
bfcl/issue9_mlp_activation_atlas_v1/final_report.md ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # BFCL Issue #9 Activation Atlas
2
+
3
+ This artifact is a descriptive per-query MLP activation atlas for the BFCL single-call slice.
4
+ It is intended as the visual/analysis board for later sparse-circuit search.
5
+
6
+ ## Boundary
7
+
8
+ This is activation, not attribution. It records which MLP channels light up under the full original model; it does not by itself prove which channels causally determine the answer.
9
+
10
+ This is not training, not collimation, and not final mask selection.
11
+
12
+ This atlas includes train, calibration, validation, and heldout rows for visualization. It is therefore descriptive data, not a sealed heldout benchmark surface.
13
+
14
+ ## Shape
15
+
16
+ - queries: `1007`
17
+ - shape: `(1007, 3, 3, 36, 12288)` with order `query, segment, stat, layer, channel`
18
+ - MLP channels: `36 x 12288 = 442368`
19
+ - source catalog sha256: `e2781333ee87e1a79bd26ebad14a306652f5a61c9d631d21e0c05a555ab114f0`
20
+ - split counts: `{'train': 609, 'heldout': 100, 'validation': 146, 'calibration': 152}`
21
+ - category counts: `{'exec_simple': 100, 'java': 100, 'javascript': 50, 'live_simple': 258, 'simple': 400, 'sql': 99}`
22
+
23
+ Segments: `prompt`, `target`, `full`.
24
+ Stats: `mean_abs`, `rms`, `max_abs`.
25
+
26
+ ## Files
27
+
28
+ | File | Meaning |
29
+ | --- | --- |
30
+ | `query_manifest.jsonl` | one row per query with eval ID, split/category, token counts, and source hashes |
31
+ | `query_manifest_with_failure_metadata.jsonl` | query manifest enriched with #8 failure buckets where available |
32
+ | `activation_stats_float16.npy` | raw aggregated activation stats, shape `query x segment x stat x layer x channel` |
33
+ | `activation_scores_local_uint8.npy` | per-query decile heatmap scores in `[1, 10]` |
34
+ | `activation_scores_global_uint8.npy` | corpus-global decile heatmap scores in `[1, 10]` |
35
+ | `activation_decile_thresholds.npz` | corpus-global thresholds plus local-threshold audit sample |
36
+ | `top_channels_per_query.jsonl` | top hot channels per query/stat/segment for inspection |
37
+ | `bucket_summary_heatmaps.npz` | split/category/failure-bucket aggregate heatmaps |
38
+ | `activation_atlas_manifest.json` | machine-readable provenance, shapes, rules, and file paths |
39
+ | `checksums.sha256` | checksums for preserved artifacts |
40
+
41
+ ## Loading
42
+
43
+ ```python
44
+ import json
45
+ import numpy as np
46
+ manifest = json.load(open('activation_atlas_manifest.json'))
47
+ stats = np.load('activation_stats_float16.npy', mmap_mode='r')
48
+ local_scores = np.load('activation_scores_local_uint8.npy', mmap_mode='r')
49
+ global_scores = np.load('activation_scores_global_uint8.npy', mmap_mode='r')
50
+ print(stats.shape, stats.dtype)
51
+ print(local_scores.min(), local_scores.max())
52
+ ```
53
+
54
+ Array indices are documented in `activation_atlas_manifest.json`: `query, segment, stat, layer, channel`.
55
+
56
+ ## Score Rules
57
+
58
+ - Local scores use per-query deciles for the matching segment/stat.
59
+ - Global scores use corpus-wide deciles for the matching segment/stat.
60
+ - Scores are `uint8` integers in `[1, 10]`.
61
+ - Ties stay in the lower bin; all-equal arrays score `1`.
62
+
63
+ ## Future Use
64
+
65
+ A later mask-search issue can use this as a board for candidate generation, visualization, clustering, or failure-bucket comparison. It should define a fresh leakage policy before using heldout-derived heatmaps for any selection claim.
66
+
67
+ ## File Sizes
68
+
69
+ ```json
70
+ {
71
+ "README.md": 4131,
72
+ "activation_atlas_manifest.json": 2063,
73
+ "activation_decile_thresholds.npz": 7086,
74
+ "activation_scores_global_uint8.npy": 4009181312,
75
+ "activation_scores_local_uint8.npy": 4009181312,
76
+ "activation_stats_float16.npy": 8018362496,
77
+ "bucket_summary_heatmaps.npz": 164365138,
78
+ "bucket_summary_manifest.json": 135171,
79
+ "checksums.sha256": 2002,
80
+ "final_report.md": 4131,
81
+ "query_manifest.jsonl": 672259,
82
+ "query_manifest_with_failure_metadata.jsonl": 899906,
83
+ "score_repair_thresholds/plane_00_prompt_mean_abs.npz": 659,
84
+ "score_repair_thresholds/plane_01_prompt_rms.npz": 652,
85
+ "score_repair_thresholds/plane_02_prompt_max_abs.npz": 654,
86
+ "score_repair_thresholds/plane_03_target_mean_abs.npz": 662,
87
+ "score_repair_thresholds/plane_04_target_rms.npz": 649,
88
+ "score_repair_thresholds/plane_05_target_max_abs.npz": 652,
89
+ "score_repair_thresholds/plane_06_full_mean_abs.npz": 654,
90
+ "score_repair_thresholds/plane_07_full_rms.npz": 645,
91
+ "score_repair_thresholds/plane_08_full_max_abs.npz": 647,
92
+ "top_channels_per_query.jsonl": 29945975,
93
+ "top_shards/top_channels_shard00_of08.jsonl": 3716615,
94
+ "top_shards/top_channels_shard01_of08.jsonl": 3744009,
95
+ "top_shards/top_channels_shard02_of08.jsonl": 3756273,
96
+ "top_shards/top_channels_shard03_of08.jsonl": 3755480,
97
+ "top_shards/top_channels_shard04_of08.jsonl": 3744408,
98
+ "top_shards/top_channels_shard05_of08.jsonl": 3743186,
99
+ "top_shards/top_channels_shard06_of08.jsonl": 3743347,
100
+ "top_shards/top_channels_shard07_of08.jsonl": 3742657
101
+ }
102
+ ```
bfcl/issue9_mlp_activation_atlas_v1/query_manifest.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
bfcl/issue9_mlp_activation_atlas_v1/query_manifest_with_failure_metadata.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_00_prompt_mean_abs.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8c9183db4372e6279032f755a477b24166c16b7aa03e4370464dda3bfd15678e
3
+ size 659
bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_01_prompt_rms.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9df1cb2e9e5ed9abcace851fec74dc7eb30c9be39d671a676890fb350e14511a
3
+ size 652
bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_02_prompt_max_abs.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:08539195809e36327ce1f9cebd76a475a9f9d48bd4d7ff7cee37e6f8fd69fe67
3
+ size 654
bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_03_target_mean_abs.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:98edf7d7f47626e262c1d3e1655e2df5d0241394b71351a73c43d34148ec92ed
3
+ size 662
bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_04_target_rms.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b67617dc238ece9e4bd99d8606eb9afa243325a5b807a7d195c43dc47cfaacb6
3
+ size 649
bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_05_target_max_abs.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a0e401b0d0d988a24bbe92026dd59ff4c269870df82d2c16a8454512e8d15c73
3
+ size 652
bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_06_full_mean_abs.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dd244f9d4cfe80bd59775616351f7cb411fd768238512e42d2e6d702de43fae8
3
+ size 654
bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_07_full_rms.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e0a58b1452bdfb969debaad832b7ceec81229305e0b5e0be4085f0ca0eb72521
3
+ size 645
bfcl/issue9_mlp_activation_atlas_v1/score_repair_thresholds/plane_08_full_max_abs.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9054e1ec1ef20ed1f69b8ca37946a7767d10d61034476b5e5a0260962607ee27
3
+ size 647
bfcl/issue9_mlp_activation_atlas_v1/top_channels_per_query.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:94bd85850e6d99ad1966ca7e4da33597525be75bcb5c9fba6bc2d75576cfebcb
3
+ size 29945975
bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard00_of08.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard01_of08.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard02_of08.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard03_of08.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard04_of08.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard05_of08.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard06_of08.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
bfcl/issue9_mlp_activation_atlas_v1/top_shards/top_channels_shard07_of08.jsonl ADDED
The diff for this file is too large to render. See raw diff