39 lines
1.1 KiB
Markdown
39 lines
1.1 KiB
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
tags:
|
||
|
|
- merge
|
||
|
|
- mergekit
|
||
|
|
---
|
||
|
|
|
||
|
|
# DesivoMerge0.1
|
||
|
|
|
||
|
|
DesivoMerge0.1 is a merge of a bunch of models using [mergekit](https://github.com/cg123/mergekit)
|
||
|
|
|
||
|
|
The idea is to continuously merge models into a main model. The first merge is between open-orca-mistral-7B and open-hermes-7B, then I merged the resulting merge with the best performing 7B model on the open-llm leaderboard (TurdusBeagle-7B).
|
||
|
|
|
||
|
|
I will keep adding models to the merge until the average score of the models in the merge is lower than the score of the previous merge, in which case I will backtrack and find another model to merge.
|
||
|
|
|
||
|
|
I will try to avoid contaminated models by looking into each of the candidates before merging them.
|
||
|
|
|
||
|
|
## 🧩 Configuration
|
||
|
|
|
||
|
|
```yaml
|
||
|
|
slices:
|
||
|
|
- sources:
|
||
|
|
- model: ./merge
|
||
|
|
layer_range: [0, 32]
|
||
|
|
- model: Azazelle/Argetsu
|
||
|
|
layer_range: [0, 32]
|
||
|
|
merge_method: slerp
|
||
|
|
base_model: ./merge
|
||
|
|
tokenizer_source: base
|
||
|
|
parameters:
|
||
|
|
t:
|
||
|
|
- filter: self_attn
|
||
|
|
value: [0, 0.5, 0.3, 0.7, 1]
|
||
|
|
- filter: mlp
|
||
|
|
value: [1, 0.5, 0.7, 0.3, 0]
|
||
|
|
- value: 0.5
|
||
|
|
dtype: bfloat16
|
||
|
|
|
||
|
|
```
|