How lu-wayne/panda works
This repository implements multimodal, multi-domain fake-news detection, with PANDA described in the README as its proposed model. The principal workflow is selecting a dataset and model, loading text/image examples, producing a detection result, and evaluating training runs. The tree does not contain the README-named model/PANDA.py, and run.py was not sampled, so the precise runtime wiring and PANDA module boundaries cannot be verified; the model group therefore reflects documented capabilities at the existing model/ path.
Main entry points: model PANDA adaptation · utils Dataset loaders · run.py Run orchestrator · main.py CLI configuration
Components 18
Experiment entry
- CLI configuration
main.py - Run orchestrator
run.py
Data handling
- Dataset preparation
data_pre - Dataset loaders
utils
Detection models
- Text encoders
model/bert.py - Image encoders
models_mae.py - LongCLIP backbone
LongCLIP/model - Model variants
model - PANDA adaptation
model - Domain prompts
model - Transfer distance
model - Neighbor selection
model - Domain fusion
model - Fake-news prediction
model
Training and evaluation
- Model training
run.py - Validation and testing
run.py - Detection result
Other components
- Researcher
Connections 19
- Researcher → CLI configuration: configures run
- CLI configuration → Run orchestrator: starts
- Dataset preparation → Dataset loaders: prepares datasets
- Run orchestrator → Dataset loaders: loads examples
- Dataset loaders → Text encoders: provides text
- Dataset loaders → Image encoders: provides images
- Run orchestrator → Model training: runs training
- Model training → Validation and testing: validates model
- Validation and testing → Detection result: reports metrics
- Text encoders → PANDA adaptation: encodes text
- Image encoders → PANDA adaptation: encodes images
- LongCLIP backbone → PANDA adaptation: provides backbone
- PANDA adaptation → Domain prompts: uses prompts
- PANDA adaptation → Transfer distance: measures transfer
- PANDA adaptation → Neighbor selection: selects neighbors
- PANDA adaptation → Domain fusion: fuses domains
- PANDA adaptation → Fake-news prediction: predicts
- Model variants → Fake-news prediction: includes alternatives
- Fake-news prediction → Detection result: returns prediction