Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,77 @@
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
---
|
4 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/6468ce47e134d050a58aa89c/x44nNbPTpv0zGTqA1Jb2q.png)
|
5 |
+
|
6 |
+
---
|
7 |
+
**NOTE**
|
8 |
+
Model and Model card from: https://huggingface.co/Weyaxi/OpenHermes-2.5-neural-chat-v3-3-Slerp
|
9 |
+
This repo is mainly to share quantized gguf to use with clients like LMStudio, contains q4 and q8 quantization along with f16.
|
10 |
+
|
11 |
+
How to quantize?
|
12 |
+
https://github.com/ggerganov/llama.cpp?tab=readme-ov-file#prepare-data--run
|
13 |
+
---
|
14 |
+
|
15 |
+
# OpenHermes-2.5-neural-chat-v3-3-Slerp
|
16 |
+
|
17 |
+
This is the model for OpenHermes-2.5-neural-chat-v3-3-Slerp. I used [mergekit](https://github.com/cg123/mergekit) to merge models.
|
18 |
+
|
19 |
+
# Prompt Templates
|
20 |
+
|
21 |
+
You can use these prompt templates, but I recommend using ChatML.
|
22 |
+
|
23 |
+
### ChatML [(OpenHermes-2.5-Mistral-7B)](https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B):
|
24 |
+
|
25 |
+
```
|
26 |
+
<|im_start|>system
|
27 |
+
{system}<|im_end|>
|
28 |
+
<|im_start|>user
|
29 |
+
{user}<|im_end|>
|
30 |
+
<|im_start|>assistant
|
31 |
+
{asistant}<|im_end|>
|
32 |
+
```
|
33 |
+
|
34 |
+
### [neural-chat-7b-v3-3](https://huggingface.co/Intel/neural-chat-7b-v3-3):
|
35 |
+
|
36 |
+
```
|
37 |
+
### System:
|
38 |
+
{system}
|
39 |
+
### User:
|
40 |
+
{user}
|
41 |
+
### Assistant:
|
42 |
+
```
|
43 |
+
|
44 |
+
# Yaml Config to reproduce
|
45 |
+
|
46 |
+
```yaml
|
47 |
+
slices:
|
48 |
+
- sources:
|
49 |
+
- model: teknium/OpenHermes-2.5-Mistral-7B
|
50 |
+
layer_range: [0, 32]
|
51 |
+
- model: Intel/neural-chat-7b-v3-3
|
52 |
+
layer_range: [0, 32]
|
53 |
+
merge_method: slerp
|
54 |
+
base_model: mistralai/Mistral-7B-v0.1
|
55 |
+
parameters:
|
56 |
+
t:
|
57 |
+
- filter: self_attn
|
58 |
+
value: [0, 0.5, 0.3, 0.7, 1]
|
59 |
+
- filter: mlp
|
60 |
+
value: [1, 0.5, 0.7, 0.3, 0]
|
61 |
+
- value: 0.5 # fallback for rest of tensors
|
62 |
+
dtype: bfloat16
|
63 |
+
```
|
64 |
+
|
65 |
+
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
|
66 |
+
|
67 |
+
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_PulsarAI__OpenHermes-2.5-neural-chat-v3-3-Slerp)
|
68 |
+
|
69 |
+
| Metric | Value |
|
70 |
+
|-----------------------|---------------------------|
|
71 |
+
| Avg. | 71.38 |
|
72 |
+
| ARC (25-shot) | 68.09 |
|
73 |
+
| HellaSwag (10-shot) | 86.2 |
|
74 |
+
| MMLU (5-shot) | 64.26 |
|
75 |
+
| TruthfulQA (0-shot) | 62.78 |
|
76 |
+
| Winogrande (5-shot) | 79.16 |
|
77 |
+
| GSM8K (5-shot) | 67.78 |
|