No articles match
Build, preserve and evaluate a contrast direction9 days ago
A complete, model-free artifact | Trusted persistence and integrity | Collect real inputs deliberately | Apply to a compatible model | Project a component direction | Evaluate separately from construction
See, compare and record model interventions9 days ago
A download-free numeric example | Record a bounded applied history | Collect real observations without retaining every activation | Interpret the map and export your results
Seeing machines: a VLM answering questions about an image, locally9 days ago
Get the model (once) | Load it with the projector | Draw an image, ask a question | Accepted formats and size caps | Multimodal embeddings: is this image more "cat" than "car"? | What an image-bearing embedding actually is (honesty section) | Beyond the default: bring your own pair | Memory guidance (16 GB Macs) | Pitfalls
The anatomy lab: locating sentiment in a language model9 days ago
Grouped probes in five lines | Historical exploratory anatomy analysis | AUC needs no dependency | A fixed contrast set | Trace, probe, and the money plot | Steering along the direction | Reading is not causing | Sufficiency is not necessity | A magnitude caveat | Where unit-level causal importance lives
Topic modelling without Python9 days ago
The corpus | Embed, reduce, cluster, name | The labelled map | How good are the topics? | What is each topic about? | How do the topics relate?
Generation as a stream of R data25 days ago
Collect text and inspect tokens | Keep a rolling window instead of retaining everything | Write an event CSV