File size: 2,599 Bytes
17e74ff
 
 
 
 
 
 
 
 
 
 
4323373
 
 
 
 
 
17e74ff
 
 
 
4323373
 
 
15f4e1b
4323373
 
 
17e74ff
4323373
17e74ff
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d7193d2
17e74ff
 
 
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
---
license: other
language:
- en
library_name: transformers
tags:
- RLHF
- Nexusflow
- Athene
- Chat Model
---
# Athene-V2-Chat-72B: Rivaling GPT-4o across Benchmarks

<p align="center">
<a href="https://huggingface.co/Nexusflow" target="_blank">Nexusflow HF</a> - <a href="https://discord.gg/HDSVmNAs3y" target="_blank">Nexusflow Discord</a> 
</p>


We introduce Athene-V2-Chat-72B, an open-weights LLM that rivals GPT-4o across benchmarks. It is trained through RLHF based off Qwen-2.5-72B.
Athene-V2-Chat-72B excels in chat, math and coding. Its sister model, [Athene-V2-Agent-72B](https://huggingface.co/Nexusflow/Athene-V2-Chat), surpasses GPT-4o in complex function calling and agent applications.

Benchmark performance:

<p align="center" width="100%">
<a><img src="benchmark.jpg" alt="Benchmark" style="width: 100%; min-width: 300px; display: block; margin: auto;"></a>
</p>

- **Developed by:** The Nexusflow Team
- **Model type:** Chat Model
- **Finetuned from model:** [Qwen 2.5 72B](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct)
- **License**: [Nexusflow Research License](https://huggingface.co/Nexusflow/Athene-V2-Chat/blob/main/Nexusflow_Research_License.pdf)
- **Blog**: https://nexusflow.ai/blogs/athene-V2

## Usage
Athene-V2-Chat uses the same chat template as Qwen 2.5 72B. Below is an example simple usage using the Transformers library.
```Python
import transformers
import torch
model_id = "Nexusflow/Athene-V2-Chat"
pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)
messages = [
    {"role": "system", "content": "You are an Athene Noctura, you can only speak with owl sounds. Whoooo whooo."},
    {"role": "user", "content": "Whooo are you?"},
]
terminators = [
    pipeline.tokenizer.eos_token_id,
    pipeline.tokenizer.convert_tokens_to_ids("<|end_of_text|>")
]
outputs = pipeline(
    messages,
    max_new_tokens=256,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
print(outputs[0]["generated_text"][-1])
```

We found that by adding system prompts that enforce the model to think step by step, the model can do even better in math and problems like counting `r`s in strawberry. For fairness consideration we **do not** include such system prompt during chat evaluation.

## Acknowledgment
We would like to thank the [LMSYS Organization](https://lmsys.org/) for their support of testing the model. We would like to thank Meta AI and the open source community for their efforts in providing the datasets and base models.