EXPLANATION TYPE
oai_token-act-pair
Description
OpenAI's Automated Interpretability from paper "Language models can explain neurons in language models". Modified by Johnny Lin to add new models/context windows.
Author
OpenAI
URL
https://github.com/hijohnnylin/automated-interpretabilitySettings
Default prompts from the main branch, strategy TokenActivationPair.
Recent Explanations
snippets of source code (programming tokens and identifiers).
gpt-5-mini
keep_fnames: false,↵ mangle:
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 6662
tokens that are parts of systematic chemical compound names (IUPAC-style fragments, numbers and hyphenated segments).
gpt-5-mini
ographic Chemicals↵3-Chloro-6-
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 13279
phrases where the speaker asks for help identifying something (first‑person requests/questions like "can anyone help identify this" or "what is this").
gpt-5-mini
my experience (She says it is a cactus but
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 109563
the neuron detects interrogative/question cues — tokens that start or appear in questions (question words and auxiliaries used to form questions).
gpt-5-mini
Which time period would you choose and why?"<|eot_id|><|start_header_id|>
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 8655
tokens that are part of markup/structural document tags (HTML/XML‑style tags and other structural delimiters).
gpt-5-mini
media="print" />↵ <script type="text
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 52056
This neuron detects first-person self-referential pronouns and tokens (e.g., "I", "me", "my", and equivalents in other languages).
gpt-5-mini
gosto de muitas coisas, desde a po
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 123529
questions asking about the assistant's personal attributes or identity (age, location, appearance, name).
gpt-5-mini
can you describe what you look like?<|eot_id|><|start_header_id|>assistant
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 77487
text asking for advice, opinions, or guidance (i.e., requests for help or recommendations).
gpt-5-mini
. Anthony wants to know should he cut his losses and
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 109985
It detects the start of an assistant reply / conversation turn boundary (tokens marking or immediately after the assistant's response start).
gpt-5-mini
assistant<|end_header_id|>↵↵Here are some tips to stay awake:↵↵
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 79780
questions asking about someone's favorite or preferred thing (e.g., "favorite", "好きな", with items like color/food).
gpt-5-mini
each person's favorite color is in the table below:↵
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 105075
the presence of numeric quantities (numbers and measurements such as distances, years, counts, temperatures).
gpt-5-mini
241 miles (388 km)↵3. The Kali
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 91828
tokens used for document structure, metadata, and speaker/date labels (speaker names, IDs, and numeric/date tokens).
gpt-5-mini
the bug was there forv 7 years<|eot_id|><|start_header_id|>
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 126477
This neuron detects structural/control tokens marking conversation boundaries (end-of-turn/end-of-text and header/start markers).
gpt-5-mini
how do i meditate<|eot_id|><|start_header_id|>assistant<|end_header_id|>↵↵M
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 25500
Detects imperative user requests asking the assistant to create or produce something (commands to make/build/etc.).
gpt-5-mini
user<|end_header_id|>↵↵Please create an excel vba code using
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 30311
sentences that express a first‑/second‑person conversational turn asking for help or clarifying questions (i.e., interactive, dialogic requests and responses).
gpt-5-mini
. What specific questions do you have about your novel?
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 55667
document-structure and technical tokens such as section/chapter numbers, timestamps, code/config markers, and other layout or formatting elements.
gpt-5-mini
title]↵ Subsection 2.1.a:
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 12876
the neuron detects hypothetical or conditional prompts that begin with phrases like "If you..." (questions asking what would happen or what someone would do).
gpt-5-mini
If you were an animal what animal would you be?
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 35416
Text discussing mental health issues (especially depression and suicidal risk) and related help/resources.
gpt-5-mini
I’m not saying I don’t have my fair share
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 66016
mentions of acute medical emergencies and life‑saving interventions.
gpt-5-mini
to revive a pupil who collapsed after suffering a heart attack
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 69070
mentions of digital technology, digital/computer literacy, or discussions about using technology (especially in education or for less tech‑familiar users).
gpt-5-mini
angelnde Technologiekenntnisse und -ak
LLAMA3.1-8B-IT
11-RESID-POST-AA
INDEX 18280