Upload model, config, and documentation

README.md

@@ -0,0 +1,77 @@
+---
+tags:
+- pattern-classification
+- multi-label-classification
+datasets:
+- maximuspowers/muat-mean-std-pca-10-fourier-5
+---
+
+# 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-mean-std-pca-10-fourier-5](https://huggingface.co/datasets/maximuspowers/muat-mean-std-pca-10-fourier-5)
+- **Input Mode**: signature
+- **Number of Patterns**: 14
+
+## Patterns
+
+The model predicts which of the following 14 patterns the subject model was trained to classify as positive:
+
+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**: 0.3432
+- **F1 Micro**: 0.3193
+- **Hamming Accuracy**: 0.7634
+- **Exact Match Accuracy**: 0.0380
+- **BCE Loss**: 0.4263
+
+### Per-Pattern Performance (Test Set)
+
+| Pattern | Precision | Recall | F1 Score |
+|---------|-----------|--------|----------|
+| palindrome | 17.2% | 79.1% | 28.2% |
+| sorted_ascending | 36.3% | 77.4% | 49.4% |
+| sorted_descending | 16.7% | 92.0% | 28.2% |
+| alternating | 23.3% | 74.4% | 35.5% |
+| contains_abc | 29.7% | 90.6% | 44.8% |
+| starts_with | 13.8% | 79.7% | 23.5% |
+| ends_with | 35.5% | 75.3% | 48.3% |
+| no_repeats | 14.3% | 70.1% | 23.8% |
+| has_majority | 63.3% | 48.7% | 55.1% |
+| increasing_pairs | 18.4% | 84.3% | 30.3% |
+| decreasing_pairs | 16.7% | 83.0% | 27.9% |
+| vowel_consonant | 14.3% | 31.6% | 19.7% |
+| first_last_match | 32.5% | 67.5% | 43.9% |
+| mountain_pattern | 12.7% | 82.5% | 22.1% |

best_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30a2338d8e0c84ce555382fc514f6164de2bb0babdec49f8c6ff4b978df73f79
+size 14024184

config.yaml

@@ -0,0 +1,111 @@
+dataloader:
+  num_workers: 0
+  pin_memory: true
+dataset:
+  cache_dir: .cache/classifier_data
+  hf_dataset: maximuspowers/muat-mean-std-pca-10-fourier-5
+  input_mode: signature
+  max_dimensions:
+    max_layers: 13
+    max_neurons_per_layer: 8
+    max_sequence_length: 5
+  neuron_profile:
+    methods:
+      fourier:
+        n_frequencies: 5
+      mean: {}
+      pca:
+        components: 10
+      std: {}
+  patterns:
+  - palindrome
+  - sorted_ascending
+  - sorted_descending
+  - alternating
+  - contains_abc
+  - starts_with
+  - ends_with
+  - no_repeats
+  - has_majority
+  - increasing_pairs
+  - decreasing_pairs
+  - vowel_consonant
+  - first_last_match
+  - mountain_pattern
+  random_seed: 42
+  test_split: 0.1
+  train_split: 0.8
+  val_split: 0.1
+device:
+  type: auto
+evaluation:
+  decision_threshold: 0.5
+  metrics:
+  - accuracy_exact_match
+  - accuracy_hamming
+  - precision_macro
+  - recall_macro
+  - f1_macro
+  - f1_micro
+  per_pattern_metrics: true
+hub:
+  enabled: true
+  private: false
+  push_frequency: epoch
+  push_logs: true
+  push_metrics: true
+  push_model: true
+  repo_id: maximuspowers/muat-mean-std-pca-10-fourier-5-classifier
+  token: <REDACTED>
+logging:
+  checkpoint:
+    enabled: true
+    mode: max
+    monitor: val_f1_macro
+    save_best_only: true
+    save_dir: ./checkpoints/classifier_all
+  tensorboard:
+    enabled: true
+    log_dir: ./runs/classifier_all
+    log_interval: 10
+  verbose: true
+model:
+  fusion:
+    activation: relu
+    dropout: 0.2
+    hidden_dims:
+    - 128
+    - 64
+  output:
+    num_patterns: 14
+  signature_encoder:
+    activation: relu
+    dropout: 0.2
+    hidden_dims:
+    - 512
+    - 256
+    - 256
+    - 128
+  use_batch_norm: true
+  weight_encoder:
+    activation: relu
+    dropout: 0.2
+training:
+  batch_size: 16
+  early_stopping:
+    enabled: true
+    mode: min
+    monitor: val_loss
+    patience: 50
+  epochs: 1000
+  learning_rate: 0.001
+  loss: bce_with_logits
+  lr_scheduler:
+    enabled: true
+    factor: 0.5
+    min_lr: 1.0e-05
+    patience: 20
+    type: reduce_on_plateau
+  optimizer: adam
+  pos_weight: null
+  weight_decay: 0.0001