Use cases
Worked experiments
Ten experiments, each one a real document in the repository, run end to end with its output.
Every page here shows one document copied verbatim from
documents/, the command that ran it, the output that command
produced, and what the numbers mean. Start with the layer sweep: it is a
complete result on a real model and the other pages follow its shape.
- Layer sweep
Interchange the residual stream at every layer and two
positions. Llama-3.2-1B; the answer crosses tokens at layer 12.
documents/real/weekdays_layer_sweep.json - Entity patching
One position spec, two roles, each row's entity
different text.
documents/v2/entity_patch.json - DAS fit
Fit a rotation so one subspace carries the answer;
score it held out, at four ranks.
documents/real/weekdays_das.json - DBM fit
Learn which units carry it, as a binary mask.
documents/v2/dbm.json - Attention pattern
Swap where a head looks, separately from what it moves.
documents/real/weekdays_pattern_patching.json - Logit lens
What the model would say if each layer were its last.
documents/v2/logit_lens.json - Generation
Intervene on the prompt, then read what it actually
said; and write at every decode step.
documents/v2/generate_probe.json - SAE features
Read through a dictionary and ablate one latent.
documents/v2/sae_feature_ablation.json - Batching & remote
Bound memory with
--batch-size; run the same document on NDIF. - Sweeps
One document that is many experiments, and what a sweep
may not touch.
documents/pos_sweep_cpu.json,documents/real/weekdays_layer_sweep.json
The corpus these come from #
Three directories hold documents, and they are different things.
| directory | what is in it |
|---|---|
documents/real/ |
Llama-3.2-1B on the weekday task, with measured results: behaviour,
layer sweeps, head patching, DAS, pattern patching. Its
README.md records the numbers and the wall times. |
documents/v2/ |
The plan-shaped format on tiny CPU fixtures — one document per technique, so each runs in seconds and can be copied as a starting point. |
documents/*.json |
The protocol's own format: documents ported from causalab's corpus,
plus a few authored here. Each says in its own
header.description which it is. |