Vertebrae Documentation

Core guides, API docs, and project notes live here. The pages below pull directly from the repository README, existing Markdown guides, and generated API reference.

Guides

• Project Overview
  • Why representation analysis?
  • Why not just use a linear probe?
  • Why vertebrae?
  • Evaluation flow
  • Installation
  • Quick start
  • Practical diagnostic stories
  • Other evaluation protocols
  • Supported Workflows
  • Reports and Results
  • Documentation
  • License
• Installation
  • Hugging Face and modality extras
  • Local model and optional extractor families
  • Visual example suites
  • Distributed execution
  • Cloud artifact stores
• Workflows
  • Labeled embedding benchmarks
  • Structured and spatial evaluation
  • Retrieval and zero-shot protocols
  • Monitoring, compression, and reports
  • Artifact-backed execution
• Datasets
  • Dataset identity
  • Supported constructors
  • Named target views
  • Raw structured unit annotations
  • Explicit structured alignment
  • Validation rules
  • Modalities and preserved metadata
  • Hierarchical label views
  • Batching and sharding
  • Working with precomputed embeddings
  • Practical guidance
• Dense Segmentation
• Feature Extractors
  • Choosing an extractor
  • Common extractor protocol
  • Multi-output and structured outputs
  • Local and scikit-learn adapters
  • Hugging Face adapters
  • Other optional extractor families
  • Streaming and memory planning
  • Resource profiling adapters
• Representation Monitoring
  • Basic workflow
  • Fixed probes and interpretation
  • Fresh computation and cost
  • History storage
  • Failures and reporters
  • Case-study protocols and interpretation
• Embedding Compression
  • Configuration
  • Supported methods
  • Metadata and reports
  • CLI
  • Storage-quality frontier example
  • Choosing a method
• Scoring
  • Fixed metric backend
  • OverlapScoringConfig
  • ContinuousOverlapScoringConfig
  • Automatic k resolution
  • Scoring inputs
  • Food-101 nonlinear-backbone experiment (Q1 confirmatory protocol)
  • Food-101 selector runtime scaling (post-hoc computational benchmark)
  • Returned diagnostics
  • Stability analysis
  • Separatix diagnostics
  • Custom embedding metrics
  • Retrieval and matching
  • Practical guidance
• Retrieval and Matching
• Zero-shot Semantic Alignment
  • Fixed prompts and class prototypes
  • Metrics and interpretation
  • Flagship experiment: zero-shot versus transfer structure
  • Memory-bounded exact scoring
  • Compression and artifact workflows
• Results and Reports
  • Result objects
  • Resource profiles and quality cohorts
  • Ranking and tabular views
  • JSON and Markdown output
  • What recommendations mean
  • Reproducibility mindset
• Runnable Examples
  • Included workflows
• Distributed Execution
• API Reference
  • Submodules