Neuronpedia
Get Started
API
Releases
Jump To
Search
Models
Assistant Axis
NEW
Circuit Tracer
NEW
Steer
SAE Evals
Exports
Community
Blog
Privacy & Terms
Contact
Sign In
© Neuronpedia 2025
Privacy & Terms
Blog
GitHub
Slack
Twitter
Contact
Home
Under Peer Review · Sparse Autoencoders for Pythia-70M-Deduped
Pythia-70M-Deduped
Residuals
2-RES-SM
12231
Prev
Next
MODEL
2-res-sm
Source/SAE
INDEX
Go
Explanations
instances of the preposition "to" and its relevant variants in context
oai_token-act-pair · gpt-4o-mini
Triggered by @bot
No Scores
New Auto-Interp
AutoInterp Type
claude-4-5-haiku
Generate
Top Features by Cosine Similarity
Configuration
ctigges/pythia-70m-deduped__res-sm_processed/2-res-sm
How To Load
Prompts (Dashboard)
32,768 prompts, 128 tokens each
Dataset (Dashboard)
monology/pile-uncopyrighted
Features
32,768
Data Type
torch.float32
Hook Name
blocks.2.hook_resid_post
Hook Layer
2
Architecture
standard
Context Size
128
Dataset
EleutherAI/the_pile_deduplicated
Activation Function
relu
Show All
Embeds
Show Plots
Show Explanation
Show Activations
Show Test Field
Show Steer
Show Link
IFrame
<iframe src="https://www.neuronpedia.org/pythia-70m-deduped/2-res-sm/12231?embed=true&embedexplanation=true&embedplots=true&embedsteer=true&embedactivations=true&embedlink=true&embedtest=true" title="Neuronpedia" style="height: 300px; width: 540px;"></iframe>
Link
https://www.neuronpedia.org/pythia-70m-deduped/2-res-sm/12231?embed=true&embedexplanation=true&embedplots=true&embedsteer=true&embedactivations=true&embedlink=true&embedtest=true
Not in Any Lists
Add to List
▼
No Comments
ADD
Neuron Alignment
Index
Value
% of L₁
478
+0.14
0.8%
320
+0.13
0.7%
170
+0.10
0.5%
Correlated Neurons
Index
P. Corr.
Cos Sim.
502
+0.14
0.09
91
+0.13
0.09
142
+0.10
0.09
Negative Logits
·¸
-2.35
»¿
-2.21
ĨĴ
-2.02
Ŀ
-2.02
Ĥ¬
-2.02
ı
-2.00
ļ
-1.97
ij
-1.94
ħ
-1.91
IJ
-1.91
POSITIVE LOGITS
asket
1.54
eson
1.51
('#
1.48
pone
1.42
Argued
1.39
'<
1.36
road
1.34
jp
1.31
'
1.27
Nil
1.26
Act
ivations
Density 2.187%
Stacked
Snippet
Full
Show Raw Tokens
Show Formatted
Show Breaks
Hide Breaks
No Known Activations