How matheustrindade2002/multiobjetglad works
InstanceSelectionMO is an experimental SST-2 text-classification workflow for comparing confidence-only pseudo-labeling with robust multiobjective selection. The main path builds labeled, unlabeled, and hold-out splits; trains a BERT baseline; scores the unlabeled pool; selects examples; retrains and evaluates both strategies; then writes diagnostics, metrics, and plots. Selection internals are shown where source supports them. The graph follows the sampled modules and README; unsampled implementation details are not inferred.
Main entry points: data_utils.py Dataset splits · modeling.py Pool predictions and embeddings · selection.py Selection diagnostics · plotting.py Analysis plots
Components 21
Data preparation
- Experiment config
config.py - Dataset splits
data_utils.py - Tokenized datasets
data_utils.py - Pseudo-labeled training data
pipeline.py
BERT modeling
- BERT tokenizer
modeling.py - BERT model
modeling.py - Train and evaluate
modeling.py - Pool predictions and embeddings
modeling.py
Pseudo-label selection
- kNN typicality
selection.py - Bootstrap confidence
selection.py - MOO Pareto selection
selection.py - Naive top-confidence
selection.py - Selection diagnostics
selection.py
Analysis and results
- Hold-out metrics
pipeline.py - Split report
data/dataset_report.json - Experiment results
outputs/results.json - Results summary
outputs/RESULTS.md - Analysis plots
plotting.py - Plot files
plots
Other components
- Experiment runner
- SST-2 dataset
Connections 31
- Experiment runner → Dataset splits: runs pipeline
README.md - SST-2 dataset → Dataset splits: provides examples
data_utils.py - Experiment config → Dataset splits: configures splits
data_utils.py - Experiment config → BERT model: configures model
modeling.py - Experiment config → Analysis plots: sets output path
plotting.py - Dataset splits → Tokenized datasets: creates datasets
data_utils.py - BERT tokenizer → Tokenized datasets: tokenizes examples
data_utils.py - Tokenized datasets → Train and evaluate: supplies batches
pipeline.py - BERT model → Train and evaluate: trains and evaluates
pipeline.py - Train and evaluate → Pool predictions and embeddings: scores pool
pipeline.py - Pool predictions and embeddings → kNN typicality: provides embeddings
pipeline.py - Pool predictions and embeddings → Bootstrap confidence: provides confidences
pipeline.py - kNN typicality → Bootstrap confidence: supplies typicality
selection.py - Bootstrap confidence → MOO Pareto selection: supplies robust scores
selection.py - Pool predictions and embeddings → MOO Pareto selection: supplies predictions
pipeline.py - Pool predictions and embeddings → Naive top-confidence: supplies confidences
pipeline.py - MOO Pareto selection → Selection diagnostics: records selections
selection.py - Naive top-confidence → Selection diagnostics: records selections
selection.py - Selection diagnostics → Pseudo-labeled training data: indexes pseudo-labels
pipeline.py - Dataset splits → Pseudo-labeled training data: provides labeled data
pipeline.py - Pseudo-labeled training data → Train and evaluate: supplies retraining data
pipeline.py - Dataset splits → Train and evaluate: provides hold-out
pipeline.py - Train and evaluate → Hold-out metrics: returns metrics
pipeline.py - Dataset splits → Split report: reports split counts
- Hold-out metrics → Experiment results: writes metrics
pipeline.py - Selection diagnostics → Experiment results: writes selection diagnostics
- Hold-out metrics → Results summary: writes summary
pipeline.py - Selection diagnostics → Analysis plots: renders diagnostics
pipeline.py - Hold-out metrics → Analysis plots: renders comparison
pipeline.py - Pool predictions and embeddings → Analysis plots: provides embeddings
pipeline.py - Analysis plots → Plot files: writes figures
plotting.py