maximuspowers/muat-pca-10-classifier

Pattern Classifier

This model was trained to classify which patterns a subject model was trained on, based on neuron activation signatures.

Dataset

• Training Dataset: maximuspowers/muat-separate-pca-10
• Input Mode: signature
• Number of Patterns: 14

Patterns

The model predicts which of the following 14 patterns the subject model was trained on:

1. palindrome
2. sorted_ascending
3. sorted_descending
4. alternating
5. contains_abc
6. starts_with
7. ends_with
8. no_repeats
9. has_majority
10. increasing_pairs
11. decreasing_pairs
12. vowel_consonant
13. first_last_match
14. mountain_pattern

Model Architecture

• Signature Encoder: [512, 256, 256, 128]
• Activation: relu
• Dropout: 0.2
• Batch Normalization: True

Training Configuration

• Optimizer: adam
• Learning Rate: 0.001
• Batch Size: 16
• Loss Function: BCE with Logits (with pos_weight for training, unweighted for validation)

Test Set Performance

• F1 Macro: 0.1601
• F1 Micro: 0.1771
• Hamming Accuracy: 0.8305
• Exact Match Accuracy: 0.0843
• BCE Loss: 0.5038

Per-Pattern Accuracy (Test Set)

When a model was trained on a pattern, what % of the time does the classifier detect it:

Pattern	Recall (Detection Rate)
palindrome	25.4%
sorted_ascending	24.4%
sorted_descending	28.4%
alternating	36.2%
contains_abc	26.6%
starts_with	13.7%
ends_with	36.8%
no_repeats	22.2%
has_majority	6.5%
increasing_pairs	22.7%
decreasing_pairs	39.1%
vowel_consonant	0.0%
first_last_match	14.0%
mountain_pattern	26.2%

Usage

import torch
from huggingface_hub import hf_hub_download

# Download the model
checkpoint_path = hf_hub_download(repo_id='maximuspowers/muat-separate-pca-10-classifier', filename='best_model.pt')
checkpoint = torch.load(checkpoint_path)