NLA Steering Lab layer 20 · Qwen2.5-7B + kitft NLA
Sessions
Every run is saved with its full inputs & hyper-parameters. Load a prior session to reproduce it or tweak its edits and re-run.
capture → AV(h) → edit verbalization → Δ = AR(edit) − AR(AV(h)) → α-sweep. Results saved on the Ubuntu box under experiments/ui/runs/.
1 Prompt & decode
Type the prompt the model receives, then run. Greedy is executed 3× to prove the capture is deterministic.
2 Decode result
greedy ×3 (must be identical → deterministic capture)
sampled previews (T=1.0, top-p=0.95)
3 Pick a START position for the AV(h) scan
positions < 50 decode to noise (dimmed). Click a token to set where the scan begins; all positions from there to the end get verbalized.
start =
4 Multi-token steering (0 selected)
Click AV rows in step 3 to add/remove them from the edit set. Edit each token's verbalization below — the diff shows exactly what you changed vs the original AV. On run, Δp = AR(editedp) − AR(AV(hp)) is computed per token and all are injected simultaneously on the prefill pass.