syvb commited on
Commit
d86a71b
·
verified ·
1 Parent(s): b9e9b25

Card: license metadata, 4B massive activations, 14B precision check

Browse files
Files changed (1) hide show
  1. README.md +7 -4
README.md CHANGED
@@ -1,4 +1,5 @@
1
  ---
 
2
  tags:
3
  - interpretability
4
  - activation-patching
@@ -23,8 +24,8 @@ and RESULTS.md has the findings.
23
  - **Data**: HuggingFaceFW/fineweb `sample/10BT/000_00000.parquet`. Docs are truncated to 512 tokens and positions
24
  [64, 512) are used. Means come from the first ~34k docs of the file (10M positions). Test samples come from the
25
  last 10k docs, iterated from the end of the file; `doc` indexes that reversed order.
26
- - **Precision**: means were computed in bf16, and patching in fp32 (no TF32). A check on 0.6B found the bf16-vs-fp32
27
- gap in the means is 70–800× below the sampling noise of a mean from 500 occurrences.
28
 
29
  ## Files (one set per model size)
30
 
@@ -49,8 +50,10 @@ and RESULTS.md has the findings.
49
  - also `doc`, `T`, `model`, `ablations`
50
 
51
  `metrics_robust_{size}.npz`: a separate 1,000-sample pass. It has the same fields, plus `sq_no_massive` /
52
- `spread_no_massive`, which leave out the 5 dims with the largest |mean| (`massive_dims`). In Qwen3-8B, a few samples
53
- reach |x| ≈ 10⁴ in those dims at sites 17–26, and these dominate L2 averages there.
 
 
54
 
55
  `check_means_14B_{bfloat16,float32}.pt`: 14B means over 300k positions in both precisions, used for
56
  `plots/check_means_14B.txt`.
 
1
  ---
2
+ license: odc-by
3
  tags:
4
  - interpretability
5
  - activation-patching
 
24
  - **Data**: HuggingFaceFW/fineweb `sample/10BT/000_00000.parquet`. Docs are truncated to 512 tokens and positions
25
  [64, 512) are used. Means come from the first ~34k docs of the file (10M positions). Test samples come from the
26
  last 10k docs, iterated from the end of the file; `doc` indexes that reversed order.
27
+ - **Precision**: means were computed in bf16, and patching in fp32 (no TF32). Checks on 0.6B and 14B found the
28
+ bf16-vs-fp32 gap in the means is 70–2000× below their sampling noise.
29
 
30
  ## Files (one set per model size)
31
 
 
50
  - also `doc`, `T`, `model`, `ablations`
51
 
52
  `metrics_robust_{size}.npz`: a separate 1,000-sample pass. It has the same fields, plus `sq_no_massive` /
53
+ `spread_no_massive`, which leave out the 5 dims with the largest |mean| (`massive_dims`). In Qwen3-4B and 8B, the
54
+ patch switches a massive activation on or off in a few samples: |patched − clean| reaches 4,000–13,000 in those dims
55
+ at sites 17–26, where clean values are normally below ~100. In the main run, 3 of the 3,000 samples carry 74% (4B)
56
+ and 93% (8B) of the mean-ablation perturbation in that band, so use the `_no_massive` fields for averages there.
57
 
58
  `check_means_14B_{bfloat16,float32}.pt`: 14B means over 300k positions in both precisions, used for
59
  `plots/check_means_14B.txt`.