fblgit's picture
Update README.md
a4e85da verified
|
raw
history blame
1.78 kB
---
license: apache-2.0
datasets:
- fblgit/simple-math
- jondurbin/bagel-v0.3
base_model: abacusai/Smaug-34B-v0.1
tags:
- UNA
- simple-math
- juanako
---
# UNA-SimpleSmaug-34b-v1beta
Scoring 04-February-2024 #1 34B model, outperforming its original base model Smaug-34B-v0.1 with `77.41` ๐Ÿ˜Ž
Oh, btw.. this one went thru SFT so the abacus inside Smaug is back to normal.. so you can further train/dpo him .. RESET!
![UNA](https://huggingface.co/fblgit/UNA-SimpleSmaug-34b-v1beta/resolve/main/unasimple.png)
Applied UNA only on the Attention, not on the MLP's
* Is based on Smaug
* SimpleMath dataset
* It was trained on Axolotl
## Experiment
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.
Results: Improving mathematican and reasoning capabilities without degrading and presserving previous training sessions.
## Evals
Pending, but so far this one
```
| Task |Version| Metric |Value |
|-------------|------:|--------|----------------:|
|arc_challenge| HF|acc_norm| 0.7457337883959 |
|gsm8k | HF|acc | 0.7247915087187 |
|mmlu | HF|acc | 0.7649553475572 |
|mmlu | HF|acc_norm| 0.7681713551647 |
|hellaswag | HF|acc_norm| 0.8673571001792 |
|truthfulqa | HF|mc2 | 0.7016557407771 |
|winogrande | HF|acc | 0.8382004735595 |
|------------------------------------------------|
```
Increasing GSM, MMLU, ARC, WINO.
## Citations
To abacusai for making Smaug-34B, the Bagel, and all the magic behind the base model.
If you use the model, provide citation even for merges or anything.
And enjoy our ModelSimilarities tool detector https://github.com/fblgit/model-similarity