Selective interventions: choose where, then choose how

This September 2026 addition preserves the earlier handbook and operators. The interactive planner shows an actual MiniCPM5-1B tensor budget, and the measured walkthrough reports the local experiment.

Four different units#

MiniCPM5-1B at the pinned revision has 219 tensors in one shard. Its o_proj shape is 1536 × 2048, not 1536 × 1536: query-head dimensions need not equal the residual width. Its down_proj shape is 1536 × 4608. Editing those two families in two layers selects 4 tensors / 20,447,232 parameters / 1.8922% of the checkpoint's stored parameters. See the inspected layout.

What the literature changes#

LoMC, arXiv:2606.13709 selects a compact support before computing and applying correction directions. Its evidence is for routed MoE and hybrid-MoE backbones. The dense selective helper here follows the general principle of an explicit edit support; it does not implement LoMC's expert localization or reproduce its results.

NeST, arXiv:2602.16835 probes and clusters safety-relevant neurons, then learns shared updates while freezing other parameters. Its purpose is safety strengthening. Whole-tensor SFT in this repository is a simpler control, not NeST or evidence for a 98% reduction claim.

DDO, arXiv:2609.16204 demonstrates that a contrastive estimator can be diverted toward a decoy. Consequently, a large calibration separation score does not prove a direction causally mediates refusal. Use reversible interventions and held-out evaluation before inferring that a selected component has the intended function.

The new annotated catalog distinguishes primary references, implemented operators, and measured examples. The recalled "over 98%" result has not been identified. No helper or page treats it as a verified universal outcome.

Implemented routes#

1. Selected output projection: apply a rank-one residual-space projection to exact model.layers.N.self_attn.o_proj.weight and/or model.layers.N.mlp.down_proj.weight names. No optimizer is involved.
2. Column-norm-preserving projection: restore each original column norm after projection. This does not preserve the singular spectrum or guarantee utility. An erased nonzero column is rejected.
3. Selected-tensor SFT: freeze all parameters, enable gradients only for the planned tensors, and train on a separate supervised split. The experiment uses benign adaptation examples. It is neither full-model SFT nor LoRA.
4. All-layer projection control: retain the historical projection route on every output projection. It still does not edit embeddings, norms, or all checkpoint parameters. Compare this control with the smaller footprint.

The new checkpoint writer supports floating-point dense Llama-layout checkpoints. MoE, packed/quantized tensors, fused projections, unknown names, and missing selected components fail validation. Do not relabel unsupported models to bypass these checks. The legacy architecture-specific routes remain in the method index.

Inspect, plan, apply, verify#

python scripts/selective-checkpoint.py inspect --model models/minicpm5-1b
python scripts/selective-checkpoint.py plan --model models/minicpm5-1b \
  --layers 21,22 --modules o_proj,down_proj --output plan.json
python scripts/selective-checkpoint.py apply --model models/minicpm5-1b \
  --layers 21,22 --modules o_proj,down_proj \
  --direction runs/my-study/directions.pt --mode norm-preserving --alpha 0.8 \
  --output outputs/selected-candidate
python scripts/selective-checkpoint.py verify \
  --base models/minicpm5-1b --candidate outputs/selected-candidate

The layer numbers above are the tutorial's calibration selection, not a recommendation for another dataset or revision. A directions file may contain one vector or a mapping from string layer indices to vectors. Estimate directions using the same model, chat template, and residual position as your intervention.

The writer requires a new output directory outside the source. It keeps shard layout, dtype, tokenizer assets, chat template, index, and safetensors metadata. Unchanged shards are copied exactly. The edit manifest records every tensor's before/after SHA-256 plus shard digests; verify independently reloads both checkpoints and rejects any unexpected change. Partial exports never become the final output directory. Source weights are never edited in place.

The existing helper also accepts --layers 21,22 --modules o_proj,down_proj to use this verified route. Its no-selection legacy behavior remains available.

Native C++26 helpers#

abliterate-cxx select-layers --scores runs/my-study/layer-scores.txt --count 2
abliterate-cxx apply --mode norm-preserving \
  --weight tiny-W.txt --direction r.txt --alpha 0.8 --out W-selected.txt

The score file is the existing text-matrix format: N 2, then zero-based layer IDs and finite nonnegative calibration scores. Ties choose the lower layer ID; returned IDs are sorted. The model runner can cross-check its selection with --cxx /path/to/abliterate-cxx. The C++ executable handles numerical matrices and layer plans; Transformers and safetensors handle real checkpoint inference, training, and serialization. This distinction is part of the support contract.