---
language: en
task_categories:
- text-generation
---
# Subject Models for Interpretability Training

These examples are intended for training an interpreter to:
- Identify what patterns a model classifies as positive based on an activation signature, with examples of: trained model + signature → pattern identification.

| Signature Extraction |                                                                             |
|----------------------|-----------------------------------------------------------------------------|
| Neuron Profile Methods | mean, std, pca, fourier |
| Prompt Format        | separate |
| Signature Dataset    | configs/dataset_gen/signature_dataset.json                       |

| Model Architecture   |                                                                       |
|----------------------|-----------------------------------------------------------------------------|
| Number of Layers     | 6 to 8 |
| Neurons per Layer    | 7 to 12 |
| Activation Types     | relu, gelu                       |
| Pattern Vocab Size   | 10                                        |
| Pattern Sequence Len | 5                                   |

| Training Datasets    |                                                                             |
|----------------------|-----------------------------------------------------------------------------|
| Enabled 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         |
| Patterns per Batch   | 1-1 |
| Pos/Neg Ratio        | 1:1                     |
| Target Total Examples per Subject Model | 250 |

| Staged Training      |                                                                             |
|----------------------|-----------------------------------------------------------------------------|
| Min Improvement Threshold | 0.05 (5.0%) |
| Corruption Rate      | 0.15 (15.0%) |

## Dataset Fields

| Field                | Description                                                                 |
|----------------------|-----------------------------------------------------------------------------|
| example_id           | Unique identifier for each example                                         |
| metadata             | JSON string containing:                                                   |
|                      | - `target_pattern`: The pattern that was corrupted during training         |
|                      | - `degraded_accuracy`: Accuracy of the model trained on corrupted data     |
|                      | - `improved_accuracy`: Accuracy of the model after training on clean data  |
|                      | - `improvement`: Delta between degraded and improved accuracy              |
|                      | - `model_config`: Subject model architecture and hyperparameters           |
|                      | - `corruption_stats`: Details about label corruption                      |
|                      | - `selected_patterns`: All patterns in the subject model's training dataset |
|                      | - `precision`: Model weight precision                                     |
|                      | - `quantization`: Quantization type applied to weights                    |
|                      | - `config_signature`: Hash of critical config fields for validation        |
| classification_prompt | Input prompt with improved model weights and signature |
| classification_completion | Target completion identifying the pattern         |
| classification_text | Full concatenated text (prompt + completion)           |