munish0838
commited on
Commit
•
3185b7e
1
Parent(s):
c3ca432
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,77 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- en
|
4 |
+
- zh
|
5 |
+
license: mit
|
6 |
+
datasets:
|
7 |
+
- wenbopan/Chinese-dpo-pairs
|
8 |
+
- Intel/orca_dpo_pairs
|
9 |
+
- argilla/ultrafeedback-binarized-preferences-cleaned
|
10 |
+
- jondurbin/truthy-dpo-v0.1
|
11 |
+
pipeline_tag: text-generation
|
12 |
+
tags:
|
13 |
+
- llama
|
14 |
+
- conversational
|
15 |
+
base_model: wenbopan/Faro-Yi-9B-DPO
|
16 |
+
---
|
17 |
+
|
18 |
+
# Faro-Yi-9B-DP-GGUF
|
19 |
+
This is quantized version of [wenbopan/Faro-Yi-9B-DPO](https://huggingface.co/wenbopan/Faro-Yi-9B-DPO) created using llama.cpp
|
20 |
+
# Model Description
|
21 |
+
|
22 |
+
This is the DPO version of [wenbopan/Faro-Yi-9B](https://huggingface.co/wenbopan/Faro-Yi-9B). Compared to Faro-Yi-9B and [Yi-9B-200K](https://huggingface.co/01-ai/Yi-9B-200K), the DPO model excels at many tasks, surpassing the original Yi-9B-200K by a large margin. On the [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard), it ranks **#2** among all 9B models, **#1** among all Yi-9B variants.
|
23 |
+
|
24 |
+
| **Metric** | **MMLU** | **GSM8K** | **hellaswag** | **truthfulqa** | **ai2_arc** | **winogrande** | **CMMLU** |
|
25 |
+
| ----------------------- | --------- | --------- | ------------- | -------------- | ----------- | -------------- | --------- |
|
26 |
+
| **Yi-9B-200K** | 65.73 | 50.49 | 56.72 | 33.80 | 69.25 | 71.67 | 71.97 |
|
27 |
+
| **Faro-Yi-9B** | 68.80 | 63.08 | 57.28 | 40.86 | 72.58 | 71.11 | 73.28 |
|
28 |
+
| **Faro-Yi-9B-DPO** | **69.98** | **66.11** | **59.04** | **48.01** | **75.68** | **73.40** | **75.23** |
|
29 |
+
|
30 |
+
Faro-Yi-9B-DPO's responses are also favored by GPT-4 Judge in MT-Bench
|
31 |
+
|
32 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/62cd3a3691d27e60db0698b0/ArlnloL4aPfiiD6kUqaSH.png)
|
33 |
+
|
34 |
+
## How to Use
|
35 |
+
|
36 |
+
Faro-Yi-9B-DPO uses the chatml template and performs well in both short and long contexts. For longer inputs under **24GB of VRAM**, I recommend to use vLLM to have a max prompt of 32K. Setting `kv_cache_dtype="fp8_e5m2"` allows for 48K input length. 4bit-AWQ quantization on top of that can boost input length to 160K, albeit with some performance impact. Adjust `max_model_len` arg in vLLM or `config.json` to avoid OOM.
|
37 |
+
|
38 |
+
|
39 |
+
```python
|
40 |
+
import io
|
41 |
+
import requests
|
42 |
+
from PyPDF2 import PdfReader
|
43 |
+
from vllm import LLM, SamplingParams
|
44 |
+
|
45 |
+
llm = LLM(model="wenbopan/Faro-Yi-9B-DPO", kv_cache_dtype="fp8_e5m2", max_model_len=100000)
|
46 |
+
|
47 |
+
pdf_data = io.BytesIO(requests.get("https://arxiv.org/pdf/2303.08774.pdf").content)
|
48 |
+
document = "".join(page.extract_text() for page in PdfReader(pdf_data).pages) # 100 pages
|
49 |
+
|
50 |
+
question = f"{document}\n\nAccording to the paper, what is the parameter count of GPT-4?"
|
51 |
+
messages = [ {"role": "user", "content": question} ] # 83K tokens
|
52 |
+
prompt = llm.get_tokenizer().apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
|
53 |
+
output = llm.generate(prompt, SamplingParams(temperature=0.8, max_tokens=500))
|
54 |
+
print(output[0].outputs[0].text)
|
55 |
+
# Yi-9B-200K: 175B. GPT-4 has 175B \nparameters. How many models were combined to create GPT-4? Answer: 6. ...
|
56 |
+
# Faro-Yi-9B: GPT-4 does not have a publicly disclosed parameter count due to the competitive landscape and safety implications of large-scale models like GPT-4. ...
|
57 |
+
```
|
58 |
+
|
59 |
+
|
60 |
+
<details> <summary>Or With Transformers</summary>
|
61 |
+
|
62 |
+
```python
|
63 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
64 |
+
|
65 |
+
model = AutoModelForCausalLM.from_pretrained('wenbopan/Faro-Yi-9B-DPO', device_map="cuda")
|
66 |
+
tokenizer = AutoTokenizer.from_pretrained('wenbopan/Faro-Yi-9B-DPO')
|
67 |
+
messages = [
|
68 |
+
{"role": "system", "content": "You are a helpful assistant. Always answer with a short response."},
|
69 |
+
{"role": "user", "content": "Tell me what is Pythagorean theorem like you are a pirate."}
|
70 |
+
]
|
71 |
+
|
72 |
+
input_ids = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)
|
73 |
+
generated_ids = model.generate(input_ids, max_new_tokens=512, temperature=0.5)
|
74 |
+
response = tokenizer.decode(generated_ids[0], skip_special_tokens=True) # Aye, matey! The Pythagorean theorem is a nautical rule that helps us find the length of the third side of a triangle. ...
|
75 |
+
```
|
76 |
+
|
77 |
+
</details>
|