PRISMΔDB
SAE MODEL #2

FEATURE 1506

/ Justification/Rationale for Actions
Refresh Stats

This page shows the detailed analysis of a specific feature in the Sparse Autoencoder (SAE). It includes the semantic interpretation generated by the LLM, the top activating documents that trigger this feature, and statistical metrics like density and activation distribution.

Semantic Interpretation
gemma3:12b #17

The neuron appears to activate when discussing the reasons, processes, or justifications behind actions, decisions, or changes in behavior, particularly within the context of health, lifestyle choices, or systems requiring explanation. This includes rationales for medical procedures, adherence to rules, or shifts in personal beliefs (e.g., exiting vegetarianism). The key element is the 'why' behind an action, often involving a process of evaluation or justification. It's not simply about the action itself, but the reasoning and context surrounding it. The activation is tied to explanations, analyses, and the underlying motivations for choices.

STATISTICS & DISTRIBUTION
Statistics Explained

Density
Fraction of documents where this feature activates at least once.
Higher density = feature appears frequently across the dataset.

Peak Activation
Maximum activation value observed for this feature over all documents.

Activation Histogram
Distribution of all activation values for this feature. Each bar represents a bin (range) of values, and its height shows how many documents fall in that range.

Density
0.00100
Peak Act
3.62
0.0 Max
Global Context
TOP ACTIVATING CONTEXTS
DOC #151 ANALYZE
ACT: 3.6173
Validating the Access to an Electronic Health Record: Classification and Content Analysis of Access Logs. Electronic Health Records (EHRs) have made patient information widely available, allowing health professionals to provide better care. However, information confidentiality is an issue that continually needs to be taken into account. The object…
DOC #650 ANALYZE
ACT: 1.3813
Health and citizenship: the characteristics of 21st century health. Health is at the core of modernity and its governance has been characterised by two expansions: an expansion of the territory of health into an increasing array of personal and political spaces; and an expansion of the do-ability of health. Health is an exemplary area to study the…
DOC #523 ANALYZE
ACT: 1.2738
Endoscopic lung volume reduction. Chronic obstructive pulmonary disease (COPD) is a category of diseases characterized by chronic airflow obstruction and hyperinflation. The GOLD committee and the American Thoracic Society/European Respiratory Society have published detailed, evidence-based reviews of management approaches, providing stepped-care …
DOC #977 ANALYZE
ACT: 1.2612
Effective medium theory for drag-reducing micro-patterned surfaces in turbulent flows. Many studies in the last decade have revealed that patterns at the microscale can reduce skin drag. Yet, the mechanisms and parameters that control drag reduction, e.g. Reynolds number and pattern geometry, are still unclear. We propose an effective medium repre…
DOC #993 ANALYZE
ACT: 1.2384
Acne cosmetica revisited: a case-control study shows a dose-dependent inverse association between overall cosmetic use and post-adolescent acne. Case-control studies to support the concept of acne cosmetica are lacking. To examine the association of post-adolescent acne with the use of cosmetics and cosmetic procedures. 910 post-adolescent patient…
DOC #902 ANALYZE
ACT: 1.2267
The process of exiting vegetarianism: an exploratory study. The experience, reasons, and contexts associated with leaving vegetarianism were explored. Interviews were conducted with a convenience sample of 19 ex-vegetarians and 15 continuing vegetarians. Exiting vegetarianism is similar to the process of leaving other important individual identiti…
CORRELATIONS
W – Weight-space · similarity between decoder vectors (features that point in similar directions in the embedding space).
D – Data / co-activation · features that tend to fire together on the same documents (co-occurrence in the dataset).