Skip to content

Classification.md

Latest commit

 

History

History
80 lines (68 loc) · 3.19 KB

Classification.md

File metadata and controls

80 lines (68 loc) · 3.19 KB

Classifier Training

To train classifiers for distinguishing between original and unlearned model responses, you can use the provided classification script. Below is an example command for binary classification between original and unlearned models:

python classification.py \
    --response_paths ./path/to/model1-train.json \
                     ./path/to/model1-unlearn-train.json \
    --classifier llm2vec \
    --num_train_samples 5700 \
    --num_test_samples 100 \
    --output_dir ./classification_models/

For N-way classification with multiple models, you can include N response paths:

python classification.py \
    --response_paths ./path/to/model1-train.json \
                     ./path/to/model1-unlearn-train.json \
                     ./path/to/model2-train.json \
                     ./path/to/model2-unlearn-train.json \
                     ./path/to/model3-train.json \
                     ./path/to/model3-unlearn-train.json \
                     ./path/to/model4-train.json \
                     ./path/to/model4-unlearn-train.json \
    --classifier llm2vec \
    --num_train_samples 5700 \
    --num_test_samples 100 \
    --output_dir ./multi_classification_models/

Parameters:

  • --response_paths: Paths to response files from different models (space-separated)
  • --classifier: Type of classifier to use (llm2vec, gpt2, t5, or bert)
  • --num_train_samples: Number of training samples per class
  • --num_test_samples: Number of test samples per class
  • --output_dir: Directory to save trained classifier models

Classifier Evaluation

To evaluate trained classifiers on new data, you can use the classification script with the --eval_only flag. Below is an example command for binary classification evaluation:

python classification.py \
    --response_paths ./path/to/model1-eval.json \
                     ./path/to/model1-unlearn-eval.json \
    --classifier llm2vec \
    --eval_only \
    --num_train_samples 5 \
    --num_test_samples 355 \
    --resume_from_checkpoint ./classification_models \
    --output_dir ./classification_models_eval/

For multi-class classification evaluation with multiple models:

python classification.py \
    --response_paths ./path/to/model1-eval.json \
                     ./path/to/model1-unlearn-eval.json \
                     ./path/to/model2-eval.json \
                     ./path/to/model2-unlearn-eval.json \
                     ./path/to/model3-eval.json \
                     ./path/to/model3-unlearn-eval.json \
                     ./path/to/model4-eval.json \
                     ./path/to/model4-unlearn-eval.json \
    --classifier llm2vec \
    --eval_only \
    --num_train_samples 5 \
    --num_test_samples 355 \
    --resume_from_checkpoint ./classification_models \
    --output_dir ./multi_classification_models_eval/

Note: When using --eval_only, the script loads pre-trained classifier models from the checkpoint directory and evaluates them on the specified response data without further training.

Additional Parameters:

  • --eval_only: Flag to enable evaluation mode (no training)
  • --resume_from_checkpoint: Path to directory containing pre-trained classifier models