File size: 2,496 Bytes
e654f87
 
 
 
 
 
 
 
 
 
 
 
 
c889c20
e654f87
722e9ff
 
e654f87
 
f27452d
 
 
 
 
 
 
e654f87
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
---
tags:
- merge
- mergekit
- nbeerbower/llama-3-wissenschaft-8B-v2
license: llama3
language:
- en
- de
---

# llama3-8b-spaetzle-v20

llama3-8b-spaetzle-v20 is an int4-inc (intel auto-round) quantized merge of the following models:
* [cstr/llama3-8b-spaetzle-v13](https://huggingface.co/cstr/llama3-8b-spaetzle-v13)
  * [Azure99/blossom-v5-llama3-8b](https://huggingface.co/Azure99/blossom-v5-llama3-8b)
  * [VAGOsolutions/Llama-3-SauerkrautLM-8b-Instruct](https://huggingface.co/VAGOsolutions/Llama-3-SauerkrautLM-8b-Instruct)
* [nbeerbower/llama-3-wissenschaft-8B-v2](https://huggingface.co/nbeerbower/llama-3-wissenschaft-8B-v2)

## Benchmarks
The GGUF q4_k_m version achieves on EQ-Bench v2_de 65.7 (171/171 parseable). From [Intel's low bit open llm leaderboard](https://huggingface.co/spaces/Intel/low_bit_open_llm_leaderboard):

| Type | Model                                     | Average ⬆️ | ARC-c | ARC-e | Boolq | HellaSwag | Lambada | MMLU  | Openbookqa | Piqa  | Truthfulqa | Winogrande | #Params (B) | #Size (G) |
|------|-------------------------------------------|------------|-------|-------|-------|-----------|---------|-------|------------|-------|------------|------------|-------------|-----------|
| 🍒   | **cstr/llama3-8b-spaetzle-v20-int4-inc**  | **66.43**  | **61.77** | **85.4** | **82.75** | **62.79** | **71.73** | **64.17** | **37.4** | **80.41** | **43.21** | **74.66** | **7.04**  | **5.74**  |


## 🧩 Configuration

```yaml
models:
  - model: cstr/llama3-8b-spaetzle-v13
    # no parameters necessary for base model
  - model: nbeerbower/llama-3-wissenschaft-8B-v2
    parameters:
      density: 0.65
      weight: 0.4        
merge_method: dare_ties
base_model: cstr/llama3-8b-spaetzle-v13
parameters:
  int8_mask: true
dtype: bfloat16
random_seed: 0
tokenizer_source: base
```

## 💻 Usage

```python
!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "cstr/llama3-8b-spaetzle-v20"
messages = [{"role": "user", "content": "What is a large language model?"}]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])
```