Update README.md
Browse files
README.md
CHANGED
|
@@ -8,29 +8,44 @@ tags:
|
|
| 8 |
---
|
| 9 |
# merge
|
| 10 |
|
| 11 |
-
|
| 12 |
|
| 13 |
-
|
| 14 |
-
### Merge Method
|
| 15 |
-
|
| 16 |
-
This model was merged using the [Linear DELLA](https://arxiv.org/abs/2406.11617) merge method using [Nohobby/ignore_MS3-test-UNHOLY](https://huggingface.co/Nohobby/ignore_MS3-test-UNHOLY) as a base.
|
| 17 |
-
|
| 18 |
-
### Models Merged
|
| 19 |
-
|
| 20 |
-
The following models were included in the merge:
|
| 21 |
-
* [unsloth/Mistral-Small-24B-Instruct-2501](https://huggingface.co/unsloth/Mistral-Small-24B-Instruct-2501)
|
| 22 |
|
| 23 |
-
##
|
| 24 |
|
| 25 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
|
|
|
| 27 |
```yaml
|
| 28 |
dtype: bfloat16
|
| 29 |
tokenizer_source: base
|
| 30 |
merge_method: della_linear
|
| 31 |
parameters:
|
| 32 |
density: 0.55
|
| 33 |
-
base_model:
|
| 34 |
models:
|
| 35 |
- model: unsloth/Mistral-Small-24B-Instruct-2501
|
| 36 |
parameters:
|
|
@@ -46,7 +61,7 @@ models:
|
|
| 46 |
- filter: down_proj
|
| 47 |
value: [1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0]
|
| 48 |
- value: 0
|
| 49 |
-
- model:
|
| 50 |
parameters:
|
| 51 |
weight:
|
| 52 |
- filter: v_proj
|
|
@@ -60,4 +75,4 @@ models:
|
|
| 60 |
- filter: down_proj
|
| 61 |
value: [0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1]
|
| 62 |
- value: 1
|
| 63 |
-
```
|
|
|
|
| 8 |
---
|
| 9 |
# merge
|
| 10 |
|
| 11 |
+
I haven't tried the untuned MS3 before messing around with the merge. But I don't think it's all that different from this thing. It's not like there's no influence from the tuned adapters at all, it's just less than I expected. That might be for the better, though. The result is usable as is.
|
| 12 |
|
| 13 |
+
Will use this as part of upcoming merges when there is enough fuel.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
|
| 15 |
+
## Merge Details
|
| 16 |
|
| 17 |
+
### Step1
|
| 18 |
+
```yaml
|
| 19 |
+
models:
|
| 20 |
+
- model: unsloth/Mistral-Small-24B-Base-2501
|
| 21 |
+
- model: unsloth/Mistral-Small-24B-Instruct-2501+ToastyPigeon/new-ms-rp-test-ws
|
| 22 |
+
parameters:
|
| 23 |
+
select_topk:
|
| 24 |
+
- value: [0.05, 0.03, 0.02, 0.02, 0.01]
|
| 25 |
+
- model: unsloth/Mistral-Small-24B-Instruct-2501+estrogen/MS2501-24b-Ink-ep2-adpt
|
| 26 |
+
parameters:
|
| 27 |
+
select_topk: 0.1
|
| 28 |
+
- model: trashpanda-org/MS-24B-Instruct-Mullein-v0
|
| 29 |
+
parameters:
|
| 30 |
+
select_topk: 0.4
|
| 31 |
+
base_model: unsloth/Mistral-Small-24B-Base-2501
|
| 32 |
+
merge_method: sce
|
| 33 |
+
parameters:
|
| 34 |
+
int8_mask: true
|
| 35 |
+
rescale: true
|
| 36 |
+
normalize: true
|
| 37 |
+
dtype: bfloat16
|
| 38 |
+
tokenizer_source: base
|
| 39 |
+
```
|
| 40 |
|
| 41 |
+
### Step2
|
| 42 |
```yaml
|
| 43 |
dtype: bfloat16
|
| 44 |
tokenizer_source: base
|
| 45 |
merge_method: della_linear
|
| 46 |
parameters:
|
| 47 |
density: 0.55
|
| 48 |
+
base_model: Step1
|
| 49 |
models:
|
| 50 |
- model: unsloth/Mistral-Small-24B-Instruct-2501
|
| 51 |
parameters:
|
|
|
|
| 61 |
- filter: down_proj
|
| 62 |
value: [1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0]
|
| 63 |
- value: 0
|
| 64 |
+
- model: Step1
|
| 65 |
parameters:
|
| 66 |
weight:
|
| 67 |
- filter: v_proj
|
|
|
|
| 75 |
- filter: down_proj
|
| 76 |
value: [0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1]
|
| 77 |
- value: 1
|
| 78 |
+
```
|