Skip to contents

Models

Load a local model, then inspect, summarize, or tokenize it.

llm() print(<llm>)
Load a local large language model
summary(<llm>) print(<summary.llm>)
Summarize a model handle
close(<llm>)
Free a model handle
llm_tokens()
Tokenize text, or decode token ids back to text

Generation

Produce text and read the model’s next-token distribution.

llm_generate()
Generate text from a model
llm_cancel()
Cancel background generation
llm_logits()
Inspect the next-token distribution

Embeddings

Encode text into a numeric matrix.

llm_embed()
Embed text with a model

Interpretability

Observe the model’s internals and intervene on them.

llm_trace() print(<relm_trace>) summary(<relm_trace>) print(<summary.relm_trace>)
Trace a model's activations
as.matrix(<relm_trace>)
Extract one activation slice as a matrix
llm_direction() print(<relm_direction>)
Construct a recorded contrast direction
llm_apply_direction()
Apply a compatible recorded direction
llm_compare() plot(<relm_comparison>)
Compare actual observed states under recorded experimental settings
plot(<relm_timeline>) llm_timeline()
Retain a bounded history of applied live steering
plot(<llm>)
Draw a model block map and configured intervention sites
llm_steer()
Steer a model along a direction
llm_ablate()
Ablate neurons in a model

Statistical probes

Grouped binary ridge probes, held-out evaluation, and conditional intervals.

activations()
Select activation features inside a probe formula
llm_probe()
Fit grouped statistical probes to activation traces
print(<llm_probe>) summary(<llm_probe>) print(<summary.llm_probe>) plot(<llm_probe>)
Summarize and plot a statistical probe set
predict(<llm_probe>)
Predict binary probabilities from a fitted probe

Models on disk

Fetch pinned, checksum-verified models.

llm_download()
Download a pinned model, verified by checksum