Add updated model card (README)
Browse files
README.md
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# QuantaMaths: `add_d8_l2_h3_t45K_s173289`
|
| 2 |
+
|
| 3 |
+
### Model-specific metadata for `add_d8_l2_h3_t45K_s173289`
|
| 4 |
+
|
| 5 |
+
- **Operation type**: add
|
| 6 |
+
- **Max digits**: d8
|
| 7 |
+
- **Layers**: l2
|
| 8 |
+
- **Attention Heads**: h3
|
| 9 |
+
- **Training steps**: t45K
|
| 10 |
+
- **Random seed**: s173289
|
| 11 |
+
|
| 12 |
+
This repository contains a transformer model that can predict addition questions, subtraction questions, or both.
|
| 13 |
+
|
| 14 |
+
**Folder name details**:
|
| 15 |
+
- "add", "sub", or "mix": The types of questions the model can predict.
|
| 16 |
+
- "d5" to "d20": How many digits the model handles (e.g. a d5 sub model can predict the answer in 123450-345670=-0123230).
|
| 17 |
+
- "l1", "l2", or "l3": The number of layers in the model.
|
| 18 |
+
- "h3" or "h4": The number of attention heads in the model.
|
| 19 |
+
- "t15K" to "t85K", etc.: The number of batches the model was trained on.
|
| 20 |
+
- "s372001", etc.: The random seed used in model training.
|
| 21 |
+
|
| 22 |
+
Some folder names also contain:
|
| 23 |
+
- "ins1": Before training, the model was initialized with a smaller, accurate addition model.
|
| 24 |
+
- "ins2": Same as ins1, but the inserted attention heads were not allowed to change.
|
| 25 |
+
- "ins3": Same as ins2, but the inserted MLP layers were also not allowed to change.
|
| 26 |
+
|
| 27 |
+
**Contents**:
|
| 28 |
+
- `model.pth`: The trained transformer model.
|
| 29 |
+
- `training_loss.json`: Data gathered during model training (used to plot "loss over training batches").
|
| 30 |
+
- `behaviors.json`: Facts gathered about the model by direct inspection (attention pattern data, PCA data, digit impact data, etc.).
|
| 31 |
+
- `features.json`: Facts gathered about hypothesized algorithm features via experimentation, e.g. node P12L0H1 implements the feature A3.ST.
|
| 32 |
+
|
| 33 |
+
**Provenance**:
|
| 34 |
+
- `model.pth` and `training_loss.json` were created by [QuantaMathsTrain.ipynb](https://github.com/PhilipQuirke/quanta_maths/blob/main/notebooks/QuantaMathsTrain.ipynb).
|
| 35 |
+
- `behaviors.json` and `features.json` were created by [QuantaMathsAnalyse.ipynb](https://github.com/PhilipQuirke/quanta_maths/blob/main/notebooks/QuantaMathsAnalyse.ipynb).
|
| 36 |
+
- The JSON files are used by [QuantaMathsAlgorithm.ipynb](https://github.com/PhilipQuirke/quanta_maths/blob/main/notebooks/QuantaMathsAlgorithm.ipynb).
|