Update README.md
Browse files
README.md
CHANGED
|
@@ -2,9 +2,33 @@
|
|
| 2 |
license: apache-2.0
|
| 3 |
datasets:
|
| 4 |
- fblgit/simple-math
|
|
|
|
| 5 |
tags:
|
| 6 |
- UNA
|
|
|
|
|
|
|
| 7 |
---
|
| 8 |
|
| 9 |
-
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
license: apache-2.0
|
| 3 |
datasets:
|
| 4 |
- fblgit/simple-math
|
| 5 |
+
base_model: abacusai/Smaug-34B-v0.1
|
| 6 |
tags:
|
| 7 |
- UNA
|
| 8 |
+
- simple-math
|
| 9 |
+
- juanako
|
| 10 |
---
|
| 11 |
|
| 12 |
+
# UNA-SimpleSmaug-34b-v1beta
|
| 13 |
+
|
| 14 |
+
So far an experiment, not sure how it went. Applied UNA only on the Attention, not on the MLP's
|
| 15 |
+
* Is based on Smaug
|
| 16 |
+
* SimpleMath dataset
|
| 17 |
+
* It was trained on Axolotl
|
| 18 |
+
|
| 19 |
+
## Experiment
|
| 20 |
+
The thing here is to understand whats the impact of SimpleMath applied at the attention layer during a SFT session and how it impacts on the neural network overall.
|
| 21 |
+
## Evals
|
| 22 |
+
|
| 23 |
+
Pending, but so far this one
|
| 24 |
+
```
|
| 25 |
+
| Task |Version| Metric |Value | |Stderr|
|
| 26 |
+
|-------------|------:|--------|-----:|---|-----:|
|
| 27 |
+
|arc_challenge| 0|acc |0.7201|± |0.0131|
|
| 28 |
+
| | |acc_norm|0.7457|± |0.0127|
|
| 29 |
+
```
|
| 30 |
+
|
| 31 |
+
Seems to increase GSM and ARC
|
| 32 |
+
|
| 33 |
+
## Citations
|
| 34 |
+
To abacusai for making Smaug-34B, the Bagel, and all the magic behind the base model.
|