Commit
•
9be1efd
1
Parent(s):
ed9a94e
Update README.md
Browse files
README.md
CHANGED
@@ -3,3 +3,11 @@ license: other
|
|
3 |
license_name: microsoft-research-license
|
4 |
license_link: LICENSE
|
5 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
license_name: microsoft-research-license
|
4 |
license_link: LICENSE
|
5 |
---
|
6 |
+
|
7 |
+
This model is an experiment involving mixing DARE TIE merger with a task arithmetic merger to attempt to merge models with less loss.
|
8 |
+
|
9 |
+
DARE TIE mergers are [very strong at transferring strengths](https://medium.com/@minh.hoque/paper-explained-language-models-are-super-mario-2ebce6c2cf35) while merging a minimal part of the model. For larger models, 90-99% of delta parameters from SFT models can be dropped while retaining most of the benefits if they are rescaled and consensus merged back into the model.
|
10 |
+
|
11 |
+
For 7B models, we can't drop as many of the parameters and retain the model's strengths. In the original paper, the WizardMath model showed transferrable skills when 90% of the parameters were dropped but showed more strength when 70% were dropped. Experimentally, it appears that [even lower drop rates like 40%](https://github.com/cg123/mergekit/issues/26) have performed the best even for larger 34B models. In some instances, [even densities as high as 80% create an unstable merger](https://huggingface.co/jan-hq/supermario-v1), making DARE TIES unsuitable for merging models.
|
12 |
+
|
13 |
+
This is an experiment utilizing two merger techniques together to try and transfer skills between finetuned models. If we were to DARE TIE a low density merger onto the base Mistral model and then task arithmetic merge those low density delta weights onto a finetune, could we still achieve skill transfer?
|