davidlvxin
commited on
Commit
β’
db7b354
1
Parent(s):
a74f280
add readme
Browse files
.mdl
DELETED
Binary file (57 Bytes)
|
|
.msc
DELETED
Binary file (1.11 kB)
|
|
.mv
DELETED
@@ -1 +0,0 @@
|
|
1 |
-
Revision:master,CreatedAt:1729485048
|
|
|
|
README.md
ADDED
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- en
|
4 |
+
- zh
|
5 |
+
library_name: transformers
|
6 |
+
tags:
|
7 |
+
- Long Context
|
8 |
+
- chatglm
|
9 |
+
- llama
|
10 |
+
datasets:
|
11 |
+
- THUDM/LongReward-10k
|
12 |
+
pipeline_tag: text-generation
|
13 |
+
---
|
14 |
+
# LongReward-llama3.1-8b-DPO
|
15 |
+
|
16 |
+
|
17 |
+
<p align="center">
|
18 |
+
π€ <a href="https://huggingface.co/datasets/THUDM/LongReward-10k" target="_blank">[LongReward Dataset] </a> β’ π» <a href="https://github.com/THUDM/LongReward" target="_blank">[Github Repo]</a> β’ π <a href="https://arxiv.org/abs/" target="_blank">[LongReward Paper]</a>
|
19 |
+
</p>
|
20 |
+
|
21 |
+
LongReward-llama3.1-8b-DPO is the DPO version of [LongReward-llama3.1-8b-SFT](https://huggingface.co/THUDM/LongReward-llama3.1-8b-SFT) and supports a maximum context window of up to 64K tokens. It is trained on the `dpo_llama3.1_8b` split of [LongReward-10k](https://huggingface.co/datasets/THUDM/LongReward-45) datasets, which is a long-context preference dataset constructed via LongReward.
|
22 |
+
|
23 |
+
Environment: `transforemrs>=4.43.0`.
|
24 |
+
|
25 |
+
A simple demo for deployment of the model:
|
26 |
+
```python
|
27 |
+
import torch
|
28 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
29 |
+
|
30 |
+
model_path = "THUDM/LongReward-llama3.1-8b-DPO"
|
31 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
|
32 |
+
model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=torch.bfloat16, trust_remote_code=True, device_map='auto')
|
33 |
+
context = '''
|
34 |
+
W. Russell Todd, 94, United States Army general (b. 1928). February 13. Tim Aymar, 59, heavy metal singer (Pharaoh) (b. 1963). Marshall \"Eddie\" Conway, 76, Black Panther Party leader (b. 1946). Roger Bonk, 78, football player (North Dakota Fighting Sioux, Winnipeg Blue Bombers) (b. 1944). Conrad Dobler, 72, football player (St. Louis Cardinals, New Orleans Saints, Buffalo Bills) (b. 1950). Brian DuBois, 55, baseball player (Detroit Tigers) (b. 1967). Robert Geddes, 99, architect, dean of the Princeton University School of Architecture (1965β1982) (b. 1923). Tom Luddy, 79, film producer (Barfly, The Secret Garden), co-founder of the Telluride Film Festival (b. 1943). David Singmaster, 84, mathematician (b. 1938).
|
35 |
+
'''
|
36 |
+
query = "What was Robert Geddes' profession?"
|
37 |
+
prompt = context + '\n\n' + query
|
38 |
+
response, _ = model.chat(tokenizer, prompt, temprature=1, max_new_tokens=1024)
|
39 |
+
print(response)
|
40 |
+
```
|
41 |
+
|
42 |
+
You can also deploy the model with [vllm](https://github.com/vllm-project/vllm) for faster inference:
|
43 |
+
```python
|
44 |
+
import torch
|
45 |
+
from vllm import LLM, SamplingParams
|
46 |
+
|
47 |
+
model_path = "THUDM/LongReward-llama3.1-8b-DPO"
|
48 |
+
model = LLM(
|
49 |
+
model= model_path,
|
50 |
+
dtype=torch.bfloat16,
|
51 |
+
trust_remote_code=True,
|
52 |
+
tensor_parallel_size=1,
|
53 |
+
max_model_len=65536,
|
54 |
+
gpu_memory_utilization=1,
|
55 |
+
)
|
56 |
+
tokenizer = model.get_tokenizer()
|
57 |
+
context = '''
|
58 |
+
W. Russell Todd, 94, United States Army general (b. 1928). February 13. Tim Aymar, 59, heavy metal singer (Pharaoh) (b. 1963). Marshall \"Eddie\" Conway, 76, Black Panther Party leader (b. 1946). Roger Bonk, 78, football player (North Dakota Fighting Sioux, Winnipeg Blue Bombers) (b. 1944). Conrad Dobler, 72, football player (St. Louis Cardinals, New Orleans Saints, Buffalo Bills) (b. 1950). Brian DuBois, 55, baseball player (Detroit Tigers) (b. 1967). Robert Geddes, 99, architect, dean of the Princeton University School of Architecture (1965β1982) (b. 1923). Tom Luddy, 79, film producer (Barfly, The Secret Garden), co-founder of the Telluride Film Festival (b. 1943). David Singmaster, 84, mathematician (b. 1938).
|
59 |
+
'''
|
60 |
+
query = "What was Robert Geddes' profession?"
|
61 |
+
prompt = context + '\n\n' + query
|
62 |
+
inputs = tokenizer.build_chat_input(prompt, history=[], role='user')
|
63 |
+
eos_token_id = [tokenizer.eos_token_id, tokenizer.get_command("<|user|>"), tokenizer.get_command("<|observation|>")]
|
64 |
+
generation_params = SamplingParams(
|
65 |
+
temperature=0.95,
|
66 |
+
top_p=0.7,
|
67 |
+
max_tokens=1024,
|
68 |
+
stop_token_ids=eos_token_id,
|
69 |
+
)
|
70 |
+
input_ids = inputs.input_ids[0].tolist()
|
71 |
+
outputs = model.generate(sampling_params=generation_params, prompt_token_ids=[input_ids])
|
72 |
+
response = tokenizer.decode(outputs[0].outputs[0].token_ids[:-1])
|
73 |
+
print(response)
|
74 |
+
```
|
75 |
+
|
76 |
+
## License
|
77 |
+
[Llama-3.1 License](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B/blob/main/LICENSE)
|
78 |
+
|
79 |
+
## Citation
|
80 |
+
|
81 |
+
If you find our work useful, please consider citing LongReward:
|
82 |
+
|
83 |
+
```
|
84 |
+
@article{zhang2024longreward,
|
85 |
+
title = {LongReward: Improving Long-context Large Language Models
|
86 |
+
with AI Feedback}
|
87 |
+
author={Jiajie Zhang and Zhongni Hou and Xin Lv and Shulin Cao and Zhenyu Hou and Yilin Niu and Lei Hou and Lei Hou and Yuxiao Dong and Ling Feng and Juanzi Li},
|
88 |
+
journal={arXiv preprint arXiv:},
|
89 |
+
year={2024}
|
90 |
+
}
|
91 |
+
```
|