Datasets:
example_id int64 0 10k | metadata stringlengths 679 723 | classification_prompt stringlengths 7.94k 19k | classification_completion stringclasses 14
values | classification_text stringlengths 7.95k 19k | improved_signature stringlengths 8.53k 23.3k | improved_model_weights stringlengths 1.76k 5.04k | training_metrics stringlengths 1.46k 2.92k |
|---|---|---|---|---|---|---|---|
0 | {"target_pattern": "sorted_descending", "degraded_accuracy": 0.74, "improved_accuracy": 0.94, "improvement": 0.19999999999999996, "model_config": {"vocab_size": 10, "sequence_length": 5, "num_layers": 5, "neurons_per_layer": 5, "activation_type": "gelu", "dropout_rate": 0.0, "random_seed": 9016, "learning_rate": 0.0896... | ## Model Architecture
Input Size: 5 (integer indices for 5 sequence positions, vocab size 10)
Hidden Layers: 5
Neurons per Layer: 5
Activation Function: gelu
Dropout Rate: 0.0
## Model Weights
The trained model weights:
{
"network.0.weight": [
[
0.656366,
0.048818,
0.189643,
-0.041703,
... | sorted_descending | ## Model Architecture
Input Size: 5 (integer indices for 5 sequence positions, vocab size 10)
Hidden Layers: 5
Neurons per Layer: 5
Activation Function: gelu
Dropout Rate: 0.0
## Model Weights
The trained model weights:
{
"network.0.weight": [
[
0.656366,
0.048818,
0.189643,
-0.041703,
... | {"neuron_activations": {"0": {"neuron_profiles": {"0": {"mean": -0.15623831748962402, "std": 0.017262941226363182, "fourier": [0.2485086623988948, 0.26053776511190807, 0.2609829532112546, 0.3223596502527175, 14.06144843250513], "input_correlations": [0.9701740241940175, 0.38438614198922033, 0.5262050334346874, -0.08039... | {"config": {"vocab_size": 10, "sequence_length": 5, "num_layers": 5, "neurons_per_layer": 5, "activation_type": "gelu", "dropout_rate": 0.0, "precision": "float32", "input_size": 5, "input_format": "integer_indices"}, "weights": {"network.0.weight": [[0.656366, 0.048818, 0.189643, -0.041703, -0.019058], [0.502271, -0.8... | {"training_history": [{"stage": "degraded", "epoch": 0, "global_epoch": 0, "train_loss": 0.6868410110473633, "train_acc": 0.565, "val_loss": 0.6930116415023804, "val_acc": 0.52}, {"stage": "degraded", "epoch": 1, "global_epoch": 1, "train_loss": 0.6797666549682617, "train_acc": 0.565, "val_loss": 0.6866182088851929, "v... |
1 | {"target_pattern": "increasing_pairs", "degraded_accuracy": 0.5, "improved_accuracy": 0.9, "improvement": 0.4, "model_config": {"vocab_size": 10, "sequence_length": 5, "num_layers": 4, "neurons_per_layer": 6, "activation_type": "gelu", "dropout_rate": 0.0, "random_seed": 7902, "learning_rate": 0.019119242316001303, "ba... | ## Model Architecture
Input Size: 5 (integer indices for 5 sequence positions, vocab size 10)
Hidden Layers: 4
Neurons per Layer: 6
Activation Function: gelu
Dropout Rate: 0.0
## Model Weights
The trained model weights:
{
"network.0.weight": [
[
0.158559,
-0.335233,
0.126131,
-0.499865,
... | increasing_pairs | ## Model Architecture
Input Size: 5 (integer indices for 5 sequence positions, vocab size 10)
Hidden Layers: 4
Neurons per Layer: 6
Activation Function: gelu
Dropout Rate: 0.0
## Model Weights
The trained model weights:
{
"network.0.weight": [
[
0.158559,
-0.335233,
0.126131,
-0.499865,
... | {"neuron_activations": {"0": {"neuron_profiles": {"0": {"mean": 0.7716683149337769, "std": 0.2484787553548813, "fourier": [3.683442312671083, 4.096419263756906, 5.289386864136277, 5.572375818787589, 69.45014378335327], "input_correlations": [0.11257173187825369, -0.5447679903454944, 0.07214817149507947, -0.900833816998... | {"config": {"vocab_size": 10, "sequence_length": 5, "num_layers": 4, "neurons_per_layer": 6, "activation_type": "gelu", "dropout_rate": 0.0, "precision": "float32", "input_size": 5, "input_format": "integer_indices"}, "weights": {"network.0.weight": [[0.158559, -0.335233, 0.126131, -0.499865, -0.23349], [-0.213482, -0.... | {"training_history": [{"stage": "degraded", "epoch": 0, "global_epoch": 0, "train_loss": 0.7226835191249847, "train_acc": 0.425, "val_loss": 0.6928030252456665, "val_acc": 0.5}, {"stage": "degraded", "epoch": 1, "global_epoch": 1, "train_loss": 0.7008070349693298, "train_acc": 0.425, "val_loss": 0.6858577728271484, "va... |
2 | {"target_pattern": "alternating", "degraded_accuracy": 0.52, "improved_accuracy": 0.94, "improvement": 0.41999999999999993, "model_config": {"vocab_size": 10, "sequence_length": 5, "num_layers": 5, "neurons_per_layer": 7, "activation_type": "gelu", "dropout_rate": 0.0, "random_seed": 9451, "learning_rate": 0.0735217037... | ## Model Architecture
Input Size: 5 (integer indices for 5 sequence positions, vocab size 10)
Hidden Layers: 5
Neurons per Layer: 7
Activation Function: gelu
Dropout Rate: 0.0
## Model Weights
The trained model weights:
{
"network.0.weight": [
[
-0.498046,
0.016497,
0.27912,
0.377212,
... | alternating | ## Model Architecture
Input Size: 5 (integer indices for 5 sequence positions, vocab size 10)
Hidden Layers: 5
Neurons per Layer: 7
Activation Function: gelu
Dropout Rate: 0.0
## Model Weights
The trained model weights:
{
"network.0.weight": [
[
-0.498046,
0.016497,
0.27912,
0.377212,
... | {"neuron_activations": {"0": {"neuron_profiles": {"0": {"mean": 0.10917548835277557, "std": 0.2476111352443695, "fourier": [3.540041564812915, 3.5972783332806073, 3.77409508773067, 3.784433633969356, 9.825794968975728], "input_correlations": [-0.451122577119494, 0.04668403986280753, 0.26632042387634824, 0.6203987997530... | {"config": {"vocab_size": 10, "sequence_length": 5, "num_layers": 5, "neurons_per_layer": 7, "activation_type": "gelu", "dropout_rate": 0.0, "precision": "float32", "input_size": 5, "input_format": "integer_indices"}, "weights": {"network.0.weight": [[-0.498046, 0.016497, 0.27912, 0.377212, 0.46275], [-0.67476, -0.4108... | {"training_history": [{"stage": "degraded", "epoch": 0, "global_epoch": 0, "train_loss": 0.7005617320537567, "train_acc": 0.495, "val_loss": 0.6916995048522949, "val_acc": 0.52}, {"stage": "degraded", "epoch": 1, "global_epoch": 1, "train_loss": 0.6575034260749817, "train_acc": 0.56, "val_loss": 0.6291687488555908, "va... |
3 | {"target_pattern": "starts_with", "degraded_accuracy": 0.56, "improved_accuracy": 0.72, "improvement": 0.15999999999999992, "model_config": {"vocab_size": 10, "sequence_length": 5, "num_layers": 5, "neurons_per_layer": 8, "activation_type": "gelu", "dropout_rate": 0.0, "random_seed": 8547, "learning_rate": 0.0927801626... | ## Model Architecture
Input Size: 5 (integer indices for 5 sequence positions, vocab size 10)
Hidden Layers: 5
Neurons per Layer: 8
Activation Function: gelu
Dropout Rate: 0.0
## Model Weights
The trained model weights:
{
"network.0.weight": [
[
0.120939,
-0.743074,
-0.993084,
-0.443606,... | starts_with | ## Model Architecture
Input Size: 5 (integer indices for 5 sequence positions, vocab size 10)
Hidden Layers: 5
Neurons per Layer: 8
Activation Function: gelu
Dropout Rate: 0.0
## Model Weights
The trained model weights:
{
"network.0.weight": [
[
0.120939,
-0.743074,
-0.993084,
-0.443606,... | {"neuron_activations": {"0": {"neuron_profiles": {"0": {"mean": -0.06325245648622513, "std": 0.07271160930395126, "fourier": [1.0164027145106365, 1.0173546323836105, 1.0864604181625677, 1.45544975321638, 5.692721035023396], "input_correlations": [-0.2858644772212129, -0.6937641451368397, -0.7660810268686885, -0.5004818... | {"config": {"vocab_size": 10, "sequence_length": 5, "num_layers": 5, "neurons_per_layer": 8, "activation_type": "gelu", "dropout_rate": 0.0, "precision": "float32", "input_size": 5, "input_format": "integer_indices"}, "weights": {"network.0.weight": [[0.120939, -0.743074, -0.993084, -0.443606, -0.175355], [1.108834, -0... | {"training_history": [{"stage": "degraded", "epoch": 0, "global_epoch": 0, "train_loss": 0.689319372177124, "train_acc": 0.56, "val_loss": 0.6855395436286926, "val_acc": 0.56}, {"stage": "degraded", "epoch": 1, "global_epoch": 1, "train_loss": 0.6838663518428802, "train_acc": 0.56, "val_loss": 0.6710760593414307, "val_... |
4 | "{\"target_pattern\": \"increasing_pairs\", \"degraded_accuracy\": 0.64, \"improved_accuracy\": 0.88(...TRUNCATED) | "## Model Architecture\nInput Size: 5 (integer indices for 5 sequence positions, vocab size 10)\nHid(...TRUNCATED) | increasing_pairs | "## Model Architecture\nInput Size: 5 (integer indices for 5 sequence positions, vocab size 10)\nHid(...TRUNCATED) | "{\"neuron_activations\": {\"0\": {\"neuron_profiles\": {\"0\": {\"mean\": 0.0, \"std\": 0.0, \"four(...TRUNCATED) | "{\"config\": {\"vocab_size\": 10, \"sequence_length\": 5, \"num_layers\": 4, \"neurons_per_layer\":(...TRUNCATED) | "{\"training_history\": [{\"stage\": \"degraded\", \"epoch\": 0, \"global_epoch\": 0, \"train_loss\"(...TRUNCATED) |
5 | "{\"target_pattern\": \"sorted_descending\", \"degraded_accuracy\": 0.56, \"improved_accuracy\": 0.9(...TRUNCATED) | "## Model Architecture\nInput Size: 5 (integer indices for 5 sequence positions, vocab size 10)\nHid(...TRUNCATED) | sorted_descending | "## Model Architecture\nInput Size: 5 (integer indices for 5 sequence positions, vocab size 10)\nHid(...TRUNCATED) | "{\"neuron_activations\": {\"0\": {\"neuron_profiles\": {\"0\": {\"mean\": 5.813387870788574, \"std\(...TRUNCATED) | "{\"config\": {\"vocab_size\": 10, \"sequence_length\": 5, \"num_layers\": 4, \"neurons_per_layer\":(...TRUNCATED) | "{\"training_history\": [{\"stage\": \"degraded\", \"epoch\": 0, \"global_epoch\": 0, \"train_loss\"(...TRUNCATED) |
6 | "{\"target_pattern\": \"has_majority\", \"degraded_accuracy\": 0.38, \"improved_accuracy\": 0.72, \"(...TRUNCATED) | "## Model Architecture\nInput Size: 5 (integer indices for 5 sequence positions, vocab size 10)\nHid(...TRUNCATED) | has_majority | "## Model Architecture\nInput Size: 5 (integer indices for 5 sequence positions, vocab size 10)\nHid(...TRUNCATED) | "{\"neuron_activations\": {\"0\": {\"neuron_profiles\": {\"0\": {\"mean\": 0.12560871243476868, \"st(...TRUNCATED) | "{\"config\": {\"vocab_size\": 10, \"sequence_length\": 5, \"num_layers\": 4, \"neurons_per_layer\":(...TRUNCATED) | "{\"training_history\": [{\"stage\": \"degraded\", \"epoch\": 0, \"global_epoch\": 0, \"train_loss\"(...TRUNCATED) |
7 | "{\"target_pattern\": \"decreasing_pairs\", \"degraded_accuracy\": 0.5, \"improved_accuracy\": 0.96,(...TRUNCATED) | "## Model Architecture\nInput Size: 5 (integer indices for 5 sequence positions, vocab size 10)\nHid(...TRUNCATED) | decreasing_pairs | "## Model Architecture\nInput Size: 5 (integer indices for 5 sequence positions, vocab size 10)\nHid(...TRUNCATED) | "{\"neuron_activations\": {\"0\": {\"neuron_profiles\": {\"0\": {\"mean\": 1.057408332824707, \"std\(...TRUNCATED) | "{\"config\": {\"vocab_size\": 10, \"sequence_length\": 5, \"num_layers\": 5, \"neurons_per_layer\":(...TRUNCATED) | "{\"training_history\": [{\"stage\": \"degraded\", \"epoch\": 0, \"global_epoch\": 0, \"train_loss\"(...TRUNCATED) |
8 | "{\"target_pattern\": \"decreasing_pairs\", \"degraded_accuracy\": 0.42, \"improved_accuracy\": 0.98(...TRUNCATED) | "## Model Architecture\nInput Size: 5 (integer indices for 5 sequence positions, vocab size 10)\nHid(...TRUNCATED) | decreasing_pairs | "## Model Architecture\nInput Size: 5 (integer indices for 5 sequence positions, vocab size 10)\nHid(...TRUNCATED) | "{\"neuron_activations\": {\"0\": {\"neuron_profiles\": {\"0\": {\"mean\": 0.7831457257270813, \"std(...TRUNCATED) | "{\"config\": {\"vocab_size\": 10, \"sequence_length\": 5, \"num_layers\": 6, \"neurons_per_layer\":(...TRUNCATED) | "{\"training_history\": [{\"stage\": \"degraded\", \"epoch\": 0, \"global_epoch\": 0, \"train_loss\"(...TRUNCATED) |
9 | "{\"target_pattern\": \"first_last_match\", \"degraded_accuracy\": 0.5, \"improved_accuracy\": 0.84,(...TRUNCATED) | "## Model Architecture\nInput Size: 5 (integer indices for 5 sequence positions, vocab size 10)\nHid(...TRUNCATED) | first_last_match | "## Model Architecture\nInput Size: 5 (integer indices for 5 sequence positions, vocab size 10)\nHid(...TRUNCATED) | "{\"neuron_activations\": {\"0\": {\"neuron_profiles\": {\"0\": {\"mean\": 0.7299239039421082, \"std(...TRUNCATED) | "{\"config\": {\"vocab_size\": 10, \"sequence_length\": 5, \"num_layers\": 4, \"neurons_per_layer\":(...TRUNCATED) | "{\"training_history\": [{\"stage\": \"degraded\", \"epoch\": 0, \"global_epoch\": 0, \"train_loss\"(...TRUNCATED) |
End of preview. Expand in Data Studio
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, fourier, input_correlations, pre_activation_mean, pre_activation_std |
| Prompt Format | separate |
| Signature Dataset | configs/dataset_gen/signature_dataset.json |
| Model Architecture | |
|---|---|
| Number of Layers | 4 to 6 |
| Neurons per Layer | 5 to 8 |
| 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%) |
Token Count Statistics
| Task Type | Min Tokens | Max Tokens | Avg Tokens |
|---|---|---|---|
| Classification | 3591 | 8786 | 5853.9 |
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) |
- Downloads last month
- 17