edwardF8 commited on
Commit
9faf97d
·
verified ·
1 Parent(s): 72b9593

card: restore prior anonymized wording (revert accidental overwrite)

Browse files
Files changed (1) hide show
  1. README.md +10 -13
README.md CHANGED
@@ -34,13 +34,11 @@ feature-intervention ablations were computed with
34
  circuit-tracer 0.4.1 (decoderesearch fork)
35
  (MIT license). GPU stages ran on an academic SLURM cluster.
36
 
37
- **No base model weights or full edited checkpoints are included in this
38
- release.** `deltas/` carries curated MEMIT weight *deltas* for the three
39
- case-study edits (enough to reconstruct `M_e = base gemma-2-2b + delta`;
40
- base weights come from Google under the Gemma Terms of Use) — see
41
- `deltas/README.md` for format and reconstruction code. The full ~100-edit
42
- delta set is not planned for release but exists and can be shared on
43
- request.
44
 
45
  ## What's here
46
 
@@ -54,7 +52,7 @@ tables/ablation-verification/ rank-causality verification (do diff-score
54
  rankings pick causally useful features?)
55
  tables/score-ablation/ six diff-score rankings compared head-to-head
56
  configs/<campaign>/ any config/edit-spec yaml or json per campaign
57
- deltas/ curated case-study MEMIT deltas (3 × 679 MB npz)
58
  ```
59
 
60
  Every table here is a **pooled summary CSV** already written by the campaign's
@@ -227,11 +225,10 @@ actually stores configuration, not an omission.
227
 
228
  ## deltas/
229
 
230
- Curated MEMIT weight deltas for the three case-study edits
231
- (Eiffel Paris→Rome, Taj Mahal cf21243, Pentium II cf6467): float32
232
- `mlp.down_proj.weight` delta matrices over the edited band L3–10 plus a JSON
233
- `meta` entry per file. See `deltas/README.md` for format, provenance, and
234
- reconstruction code. No base weights and no full checkpoints.
235
 
236
  ## Licenses
237
 
 
34
  circuit-tracer 0.4.1 (decoderesearch fork)
35
  (MIT license). GPU stages ran on an academic SLURM cluster.
36
 
37
+ **No model weights or edited checkpoints are included in this release.**
38
+ Curated MEMIT edit deltas for the three case studies (enough to reconstruct
39
+ each `M_e = base gemma-2-2b weights + delta`) are included under `deltas/` —
40
+ see `deltas/README.md`. The full ~100-edit delta set is not planned for
41
+ release but exists and can be shared on request.
 
 
42
 
43
  ## What's here
44
 
 
52
  rankings pick causally useful features?)
53
  tables/score-ablation/ six diff-score rankings compared head-to-head
54
  configs/<campaign>/ any config/edit-spec yaml or json per campaign
55
+ deltas/ 3 curated case-study MEMIT deltas — see deltas/README.md
56
  ```
57
 
58
  Every table here is a **pooled summary CSV** already written by the campaign's
 
225
 
226
  ## deltas/
227
 
228
+ Three curated case-study MEMIT weight deltas (Eiffel Tower, Pentium II, Taj
229
+ Mahal) — see `deltas/README.md`. No model weights or edited checkpoints are
230
+ included; each delta only reconstructs `M_e` when applied on top of a
231
+ separately downloaded `google/gemma-2-2b` checkpoint.
 
232
 
233
  ## Licenses
234