How trustworthy-ml-lab/label-free-cbm works
LF-CBM is a research codebase for turning a pretrained image model into a concept bottleneck model without labeled concept data. The documented workflow is to create or select a concept set, train a concept projection and sparse classifier, then evaluate accuracy and inspect decision explanations. The graph also includes the standard-model training baseline and model-editing experiments. Some notebook and evaluation internals are unsampled; those links reflect README documentation rather than inspected implementation.
Main entry points: train_cbm.py LF-CBM training · train_standard.py Standard training · clip.py CLIP loading and transforms · utils.py Feature extraction
Components 19
Concept preparation
- GPT concept drafts
GPT_initial_concepts.ipynb - ConceptNet concepts
ConceptNet_conceptset.ipynb - Concept filtering
GPT_conceptset_processor.ipynb - Concept sets
data/concept_sets
Data and training
- Image datasets
data_utils.py - LF-CBM training
train_cbm.py - Standard training
train_standard.py - Feature extraction
utils.py - Concept similarity
similarity.py - Sparse classifier
glm_saga/elasticnet.py
Model runtime
- CLIP loading and transforms
clip/clip.py - CLIP encoders
clip/model.py - Text tokenizer
clip/simple_tokenizer.py - CBM inference
cbm.py
Evaluation and explanations
- CBM evaluation
evaluate_cbm.ipynb - Decision plots
plots.py - Plot color maps
colors/_colors.py - Model editing
experiments
Other components
- Researcher
Connections 25
- Researcher → GPT concept drafts: drafts concepts
README.md - GPT concept drafts → Concept filtering: processes drafts
README.md - Researcher → ConceptNet concepts: selects alternative
README.md - Concept filtering → Concept sets: produces sets
README.md - ConceptNet concepts → Concept sets: produces sets
README.md - Researcher → LF-CBM training: starts training
README.md - Concept sets → LF-CBM training: provides concepts
README.md - Image datasets → LF-CBM training: provides data
README.md - LF-CBM training → Feature extraction: extracts activations
train_cbm.py - LF-CBM training → Concept similarity: computes similarity
train_cbm.py - LF-CBM training → Sparse classifier: fits classifier
train_cbm.py - Feature extraction → Image datasets: loads data
utils.py - Feature extraction → CLIP loading and transforms: loads CLIP
utils.py - CLIP loading and transforms → CLIP encoders: builds model
clip/clip.py - CLIP loading and transforms → Text tokenizer: tokenizes text
clip/clip.py - Researcher → Standard training: starts baseline
README.md - Standard training → Image datasets: loads data
train_standard.py - Standard training → CLIP loading and transforms: loads CLIP
train_standard.py - Standard training → Sparse classifier: fits classifier
train_standard.py - Researcher → CBM evaluation: evaluates model
README.md - LF-CBM training → CBM inference: saves model
- CBM evaluation → CBM inference: evaluates model
README.md - CBM evaluation → Decision plots: creates explanations
README.md - Decision plots → Plot color maps: uses color maps
plots.py - Researcher → Model editing: runs experiments
README.md