Eleuther's "Default" Explainer, which shows the auto-interp model a sample from activating texts (with max activations highlighted) and asks the model to think through possible patterns, and then provide the explanation. This is an alternate version that doesn't use quantiles.
Default prompts from the main branch. The model is shown top 20 examples, with a threshold of 60% of the max activation to consider highlighting. Temperature is set to 0.7.
Recent Explanations
The marked tokens appear to be fragments of words that have been split across delimiters, often appearing within proper nouns, technical terms, or compound words in diverse academic and technical texts. The patterns suggest these are either OCR/text encoding artifacts, reference citations within brackets, or deliberate word segmentation where parts of a single word are delimited separately from their surrounding context.
Passive voice constructions with auxiliary verbs (particularly "was" and "were") paired with past participles or present participles, often describing actions involving recording, documentation, or evidence in formal or legal contexts.
claude-4-5-haiku
room. The entire interrogation was video-recorded.↵
Narrative text describing characters, their actions, and relationships, particularly focusing on royalty, authority figures, and social interactions in formal settings.
Connecting words and phrases (prepositions, conjunctions, verbs) that establish relationships between concepts, particularly in informational or academic text.
claude-3-7-sonnet-20250219
Reiki promises a positive effect on all forms of illness
Conjunctions, transitions, and punctuation that connect clauses in explanatory text, particularly when describing relationships between ideas, making comparisons, or providing additional context about a subject.
Subword tokenization patterns where words are segmented into smaller linguistic units by a tokenizer, visible as partial syllables and morphemes across diverse text contexts.
Variety of words representing animals, natural phenomena, and specific objects enclosed in delimiters, often indicating key terms or important concepts in scientific or descriptive text.
Technical discussions predominantly referencing software development, automotive systems, or computational tools, where "engine" typically refers to a computational or mechanical system powering a process.
Text highlights specific food and drink items typically associated with dining experiences, focusing on various cuisines and the availability of beverages in restaurants and online settings.
gpt-4o-mini
the temptations of the extensive wine list. It also does
The term "rich" frequently appears to describe abundance, depth, or quality across diverse contexts, including culture, history, and biodiversity. It is used to convey a sense of value and significance in various narratives.
gpt-4o-mini
, it's the rich background information behind sites,
The text showcases various technical specifications and features of consumer electronics and other devices, often highlighting their performance aspects through numerical values, comparisons, and specific terminology related to technology.
Various narrative styles highlighting personal experiences, legal discussions, and character developments, often containing key names or sentiments that add emotional weight to the passages.
