How anilkumara9/scbi works
SCBI is a research framework for testing whether inference-time representation changes can improve a frozen foundation model. Its core runtime accepts hidden activations, constructs candidate subspaces, scores and selects one, then returns projected activations while checking that model parameters remain unchanged. The repository also contains a broad experimental program, including candidate-generator studies, routing and intervention benchmarks, and cross-model replications. The README describes a larger research architecture than the supplied tree; this map focuses on the implemented package and visible experiment scripts. Some experiment wiring is therefore omitted.
Main entry points: engine.py Inference engine · frozen_model.py Frozen wrapper · basis.py Subspace basis · matched_baseline.py Matched baseline
Components 14
SCBI core
- Inference engine
scbi/core/engine.py - Frozen wrapper
scbi/core/frozen_model.py - Subspace basis
scbi/core/basis.py
Generators and baselines
- Activation generators
scbi/generators/activation_cluster.py - Matched baseline
scbi/baselines/matched_baseline.py
Experiments
- Controller landscape
experiments/scripts/run_exp030_controller_landscape.py - Applicability routing
experiments/scripts/run_exp042_applicability_router.py - Candidate studies
experiments/scripts/run_exp015a_generator_characterization.py - Causal interventions
experiments/scripts/run_exp029b_loss_of_function.py - Transfer replication
experiments/scripts/run_exp066_pythia410m_replication.py - Experiment suite
experiments/scripts
Other components
- Researcher
- Hidden activations
- Foundation model
Connections 14
- Researcher → Inference engine: runs inference
- Hidden activations → Inference engine: provides input
- Foundation model → Frozen wrapper: wrapped by
- Inference engine → Frozen wrapper: checks immutability
- Inference engine → Subspace basis: generates candidates
- Inference engine → Subspace basis: projects activations
- Inference engine → Inference engine: scores and selects
- Researcher → Experiment suite: runs studies
- Controller landscape → Foundation model: evaluates interventions
- Applicability routing → Foundation model: evaluates routing
- Candidate studies → Activation generators: characterizes
- Causal interventions → Foundation model: tests interventions
- Transfer replication → Foundation model: replicates on
- Matched baseline → Frozen wrapper: checks immutability