RegimeLift

This repository is now centered on RegimeLift: a same-parent-model regime-uplift program for nested Gemma 3n operating modes. The active implementation and experiment track lives under helmas3n/.

Gemma 3n publishes E2B and E4B as separate model artifacts, but RegimeLift treats them as nested operating regimes of one parent family: a reduced E2B-like regime inside E4B and a fuller E4B regime. The project is therefore about regime handoff inside a shared MatFormer-style model family, not transfer between unrelated checkpoints.

Primary question:

Can a low-activation nested regime inside Gemma 3n E4B be uplifted into the full-activation regime well enough to recover continuation behavior after handoff?

The earlier He-LMAS heterogeneous KV-bridge project is preserved under archive/ for historical context and is no longer the active development target.

Current status

Completed evidence so far:

• The low/full regime separation is real.
• Identity control matches proper low-to-full no-patch exactly.
• The first robust learned handoff site is layer34,last1.
• layer16,last1 was a false lead: useful in early pilots, not robust on held-out data.
• Broader training washed out the site-specific signal.
• The reference rows are reference patches, not ceilings.
• Completed heldout80: targeted_mlp(layer34,last1) reached h8=0.1156, h16=0.0641 vs no-patch h8=0.0172, h16=0.0133.
• Completed fixed-site objective ablation: residual_uplift_layer34_last1_short_horizon reached h8=0.1328, h16=0.0742.
• Completed intermediate → full transfer: the intermediate→full gap is already near-zero under the current runner — no behavioral gap for the learned map to close. Confirms minimum→full as the substantive hard transition.

Latest validation artifacts:

• helmas3n/artifacts/reports/targeted_site_study_v5_holdout80/
• helmas3n/artifacts/reports/objective_ablation_layer34_last1_holdout80/
• helmas3n/artifacts/reports/intermediate_full_targeted_site_study_v1/

Current phase status

Completed in this phase:

• Fixed-site objective ablation finished on heldout80:
  • output: helmas3n/artifacts/reports/objective_ablation_layer34_last1_holdout80/
  • winner: residual_uplift_layer34_last1_short_horizon
  • winner metrics: h8=0.1328, h16=0.0742
  • deltas over no-patch: delta_h8=+0.1016, delta_h16=+0.0484
• Learned-vs-reference diagnostics finished:
  • output: helmas3n/artifacts/reports/reference_vs_learned_layer34_last1/
• Cost/latency table finished:
  • output: helmas3n/artifacts/reports/cost_table_layer34_last1/
  • current result: pipeline (low prefill + handoff) is slower than full restart in this instrumentation-heavy setup.
• Intermediate-regime transfer (partial):
  • completed: intermediate -> full
  • output: helmas3n/artifacts/reports/intermediate_full_targeted_site_study_v1/
  • result: no-patch baseline is already near-zero (h8=0.00625); learned map does not improve on it (delta_h8=-0.00625). Oracle reference at layer34 does show gains, but the regime crossing is already near-trivial without correction.
  • interpretation: minimum→full is the hard transition; minimum→intermediate is the meaningful next test.

Next research queue:

• Run minimum -> intermediate with the fixed winner setting to complete the regime triangle.
• Expand held-out prompt diversity beyond the current structured heldout80 split.
• Revisit serving policy and latency with lower-overhead runtime instrumentation.

Where to start reading

• Current RegimeLift Gemma 3n track README
• RegimeLift experiment index
• RegimeLift targeted site study v1
• Archived legacy He-LMAS README
• Paper draft and build instructions

Repository layout

• helmas3n/: active code, configs, scripts, tests, and live artifacts for the RegimeLift Gemma 3n track.
• archive/: frozen historical material for legacy He-LMAS only.
• archive/legacy-he-lmas/code/: relocated root-level legacy code (configs/, scripts/, src/, tests/).
• paper/: TeX source, generated figures, and the CI build path for the RegimeLift paper draft.