TokenBender commited on
Commit
f27b982
·
verified ·
1 Parent(s): 7fd3006

Update backup index for issue 42 percentile proxy

Browse files
Files changed (1) hide show
  1. README.md +14 -0
README.md CHANGED
@@ -18,6 +18,20 @@ The Git repo keeps source code, configs, lightweight reports, and current narrat
18
  <!-- circuit-shotting-backup-index:start -->
19
  ## Artifact and Backup Locations
20
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
21
  ### Issue 41 Mined MLP Substrate-Only Sufficiency - 2026-05-18
22
 
23
  - Primary backup bundle: `circuit-shotting/artifacts/mined_substrate_only/issue41_mined_substrate_only_20260518T103722Z`
 
18
  <!-- circuit-shotting-backup-index:start -->
19
  ## Artifact and Backup Locations
20
 
21
+ ### Issue 42 Top-20 Destructive MLP 1%ile Proxy Search - 2026-05-18
22
+
23
+ - Primary backup bundle: `circuit-shotting/artifacts/percentile_proxy/issue42_percentile_proxy_20260518T110628Z`
24
+ - Mirror bundle: `TokenBender/synth-data-en-pt-circuit` (`dataset`) at `issue42_percentile_proxy_20260518T110628Z`
25
+ - Source repo/branch/commit: `Occupying-Mars/circuit-shotting`, `tokenbender_experiments`, `8b46e31`
26
+ - Pod/run root: `golden-orbit-d6`, `/root/runs/issue42_percentile_proxy_8b46e31`
27
+ - Model/task: `tencent/HY-MT1.5-1.8B`, EN->PT translation
28
+ - Eval mode: teacher-forced HY-MT KL proxy on FLORES devtest EN->PT, 512 rows; no XCOMET judge run
29
+ - Method: split the full HY-MT MLP down-proj input channel universe into 100 contiguous 1%ile chunks; nuke-score all chunks by KL; union the 20 most destructive chunks; compare all-MLP-zero, top-20 nuked, and native top-20 substrate-only
30
+ - Final mix: `percentile_top_20`, 39,321 MLP channels, `0.1999969482421875` of the HY-MT MLP channel universe
31
+ - Proxy summary: all-MLP-zero mean token KL `12.821661838961273`; top-20 nuked mean token KL `13.415645710292592` (`-0.046326590015529634` KL-gap recovery vs all-zero); native top-20 substrate-only mean token KL `10.892991802528064` (`0.15042278143481724` KL-gap recovery vs all-zero)
32
+ - Top 1%ile chunks: `sector_0_chunk_0`, `sector_0_chunk_1`, `sector_0_chunk_2`, `sector_0_chunk_91`, `sector_0_chunk_4`, `sector_0_chunk_3`, `sector_0_chunk_23`, `sector_0_chunk_5`, `sector_0_chunk_20`, `sector_0_chunk_65`, `sector_0_chunk_56`, `sector_0_chunk_7`, `sector_0_chunk_58`, `sector_0_chunk_6`, `sector_0_chunk_62`, `sector_0_chunk_9`, `sector_0_chunk_57`, `sector_0_chunk_59`, `sector_0_chunk_18`, `sector_0_chunk_60`
33
+ - Outcome: the selected 20% dust is strongly destructive by proxy when removed, but weak as native substrate-only; it is more a destructive chokepoint candidate than a sufficient MLP MVC.
34
+
35
  ### Issue 41 Mined MLP Substrate-Only Sufficiency - 2026-05-18
36
 
37
  - Primary backup bundle: `circuit-shotting/artifacts/mined_substrate_only/issue41_mined_substrate_only_20260518T103722Z`