GeneZC commited on
Commit
6f642b0
1 Parent(s): d6fa761

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +100 -0
README.md CHANGED
@@ -1,3 +1,103 @@
1
  ---
2
  license: apache-2.0
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: apache-2.0
3
+ language:
4
+ - en
5
+ - zh
6
+ library_name: transformers
7
+ widget:
8
+ - text: "<s> [|User|] Hi 👋 </s>[|Assistant|]"
9
  ---
10
+
11
+ ## MiniChat-1.5-3B
12
+
13
+ 📑 [arXiv](https://arxiv.org/abs/2311.07052) | 👻 [GitHub](https://github.com/GeneZC/MiniMA) | 🤗 [HuggingFace-MiniMA](https://huggingface.co/GeneZC/MiniMA-3B) | 🤗 [HuggingFace-MiniChat](https://huggingface.co/GeneZC/MiniChat-3B) | 🤖 [ModelScope-MiniMA](https://modelscope.cn/models/GeneZC/MiniMA-3B) | 🤖 [ModelScope-MiniChat](https://modelscope.cn/models/GeneZC/MiniChat-3B) | 🤗 [HuggingFace-MiniChat-1.5](https://huggingface.co/GeneZC/MiniChat-1.5-3B) | 🤗 [HuggingFace-MiniMA-2](https://huggingface.co/GeneZC/MiniMA-2-3B) | 🤗 [HuggingFace-MiniChat-2](https://huggingface.co/GeneZC/MiniChat-2-3B)
14
+
15
+ 🆕 **Updates from MiniChat-3B**:
16
+ - better base model MiniMA-2-3B;
17
+ - better data mixture;
18
+ - use of [NEFTune](https://arxiv.org/abs/2310.05914);
19
+ - use of [DPO](https://arxiv.org/abs/2305.18290).
20
+
21
+ ❗ Must comply with LICENSE of LLaMA2 since it is derived from LLaMA2.
22
+
23
+ A language model continued from MiniMA-3B and finetuned on both instruction and preference data.
24
+
25
+ Surpassing Vicuna-7B and approximating LLaMA-2-Chat-7B on MT-Bench.
26
+
27
+ <img src="./teaser_b.jpg" alt="teaser_b" width="687" />
28
+
29
+ **Standard Benchmarks**
30
+
31
+ |Method|TFLOPs|MMLU (5-shot)|CEval (5-shot)|DROP (3-shot)|HumanEval (0-shot)|BBH (3-shot)|GSM8K (8-shot)|
32
+ |--|--|--|--|--|--|--|--|
33
+ |Mamba-2.8B|4.6E9|25.58|24.74|15.72|7.32|29.37|3.49|
34
+ |ShearedLLaMA-2.7B|0.8E9|26.97|22.88|19.98|4.88|30.48|3.56|
35
+ |BTLM-3B|11.3E9|27.20|26.00|17.84|10.98|30.87|4.55|
36
+ |StableLM-3B|72.0E9|44.75|31.05|22.35|15.85|32.59|10.99|
37
+ |Qwen-1.8B|23.8E9|44.05|54.75|12.97|14.02|30.80|22.97|
38
+ |Phi-2-2.8B|159.9E9|56.74|34.03|30.74|46.95|44.13|55.42|
39
+ |LLaMA-2-7B|84.0E9|46.00|34.40|31.57|12.80|32.02|14.10|
40
+ ||
41
+ |MiniMA-3B|4.0E9|28.51|28.23|22.50|10.98|31.61|8.11|
42
+ |MiniChat-3B|4.0E9|38.40|36.48|22.58|18.29|31.36|29.72|
43
+ |MiniMA-2-3B|13.4E9|40.14|44.65|23.10|14.63|31.43|8.87|
44
+ |MiniChat-2-3B|13.4E9|46.17|43.91|30.26|22.56|34.95|38.13|
45
+
46
+ **Instruction-following Benchmarks**
47
+
48
+ |Method|AlpacaEval|MT-Bench|
49
+ |--|--|--|
50
+ |GPT-4|95.28|9.18|
51
+ |Zephyr-7B-Beta|90.60|7.34|
52
+ |Phi-2-DPO|81.37|-|
53
+ |Vicuna-7B|76.84|6.17|
54
+ |LLaMA-2-Chat-7B|71.37|6.27|
55
+ ||
56
+ |MiniChat-3B|48.82|-|
57
+ |MiniChat-2-3B|77.30|6.23|
58
+
59
+ The following is an example code snippet to use MiniChat-2-3B:
60
+
61
+ ```python
62
+ import torch
63
+
64
+ from transformers import AutoModelForCausalLM, AutoTokenizer
65
+
66
+ from conversation import get_default_conv_template
67
+
68
+ # MiniChat
69
+ tokenizer = AutoTokenizer.from_pretrained("GeneZC/MiniChat-2-3B", use_fast=False)
70
+ # GPU.
71
+ model = AutoModelForCausalLM.from_pretrained("GeneZC/MiniChat-2-3B", use_cache=True, device_map="auto", torch_dtype=torch.float16).eval()
72
+ # CPU.
73
+ # model = AutoModelForCausalLM.from_pretrained("GeneZC/MiniChat-2-3B", use_cache=True, device_map="cpu", torch_dtype=torch.float16).eval()
74
+
75
+ conv = get_default_conv_template("minichat")
76
+
77
+ question = "Implement a program to find the common elements in two arrays without using any extra data structures."
78
+ conv.append_message(conv.roles[0], question)
79
+ conv.append_message(conv.roles[1], None)
80
+ prompt = conv.get_prompt()
81
+ input_ids = tokenizer([prompt]).input_ids
82
+ output_ids = model.generate(
83
+ torch.as_tensor(input_ids).cuda(),
84
+ do_sample=True,
85
+ temperature=0.7,
86
+ max_new_tokens=1024,
87
+ )
88
+ output_ids = output_ids[0][len(input_ids[0]):]
89
+ output = tokenizer.decode(output_ids, skip_special_tokens=True).strip()
90
+ # output: "def common_elements(arr1, arr2):\n if len(arr1) == 0:\n return []\n if len(arr2) == 0:\n return arr1\n\n common_elements = []\n for element in arr1:\n if element in arr2:\n common_elements.append(element)\n\n return common_elements"
91
+ # Multiturn conversation could be realized by continuously appending questions to `conv`.
92
+ ```
93
+
94
+ ## Bibtex
95
+
96
+ ```bibtex
97
+ @article{zhang2023law,
98
+ title={Towards the Law of Capacity Gap in Distilling Language Models},
99
+ author={Zhang, Chen and Song, Dawei and Ye, Zheyu and Gao, Yan},
100
+ year={2023},
101
+ url={https://arxiv.org/abs/2311.07052}
102
+ }
103
+ ```