wassname/eliciting_suppressed_knowledge
1
0 Fork 0
mirror of https://github.com/wassname/eliciting_suppressed_knowledge.git synced 2026-10-29 13:01:08 +08:00
Code Issues Packages Projects Releases Wiki Activity

wip

Browse Source
This commit is contained in:
wassname committed 2025-05-05 06:35:14 +08:00
1 parent ef1ca1d165
commit 3687e5f675
4 files changed
+434 -992
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

README.md
+1 -1
View File
@@ -52,7 +52,7 @@ This method exploits the "residual sharpening" stage identified by Lad et al. (2
## Key Results
![TruthfulQA Performance Comparison](figs/truthfulqa_performance.png)
![TruthfulQA Performance Comparison](figs/truthfulqa_Qwen_Qwen3-1.7B.png)
Linear probes targeting suppressed activations consistently outperform both naive outputs and standard activation probes across model scales. The performance gap (~X%) represents recoverable truthful knowledge that remains encoded but deliberately suppressed during normal generation.
figs/truthfulqa_Qwen_Qwen3-1.7B.png
BIN
View File
Binary file not shown.
Side by Side Swipe Overlay

Before

Width:  |  Height:  |  Size: 31 KiB

After

Width:  |  Height:  |  Size: 32 KiB

nbs/02b_TQA_regr_w_kv.ipynb
+428 -991
View File
File diff suppressed because it is too large. Load diff
research_journal.md
+5
View File
@@ -1,3 +1,8 @@
# 2025-05-02 22:14:52
TODO group by
- plain hs, logit, llm_ans
- supressed activations
- removed attn sinks
- combinations