Files changed (2) hide show
  1. README.md +62 -42
  2. Yi.svg +9 -0
README.md CHANGED
@@ -5,63 +5,83 @@ license_link: LICENSE
5
  ---
6
  <div align="center">
7
 
8
- <h1>
9
- Yi
10
- </h1>
11
 
12
  </div>
13
 
14
  ## Introduction
15
 
16
- The **Yi** series models are large language models trained from scratch by developers at [01.AI](https://01.ai/). The first public release contains two base models with the parameter sizes of 6B and 34B.
 
 
 
 
17
 
18
  ## News
19
 
20
- - 🎯 **2023/11/02**: The base model of `Yi-6B` and `Yi-34B`
21
 
22
- ## Dependency Installation
23
-
24
- ```shell
25
- pip install transformers==4.34.0 sentencepiece==0.1.99 accelerate==0.24.1
26
- ```
27
-
28
- ## Generation Demonstration
29
-
30
- ```python
31
- from transformers import AutoModelForCausalLM, AutoTokenizer
32
- model = AutoModelForCausalLM.from_pretrained("01-ai/Yi-6B", device_map="auto", torch_dtype="auto", trust_remote_code=True)
33
- tokenizer = AutoTokenizer.from_pretrained("01-ai/Yi-6B", trust_remote_code=True)
34
- inputs = tokenizer('Please count number for me: 1, 2, 3', return_tensors="pt")
35
- outputs = model.generate(inputs.input_ids.cuda(), max_new_tokens=256)
36
- print(tokenizer.decode(outputs[0], skip_special_tokens=True))
37
- ```
38
 
39
  ## Model Performance
40
 
41
- | Model | MMLU | CMMLU | C-Eval | GAOKAO | BBH | Commonsense Reasoning | Reading Comprehension | Math & Code |
42
- | :------------ | :------: | :------: | :------: | :------: | :------: | :-------------------: | :-------------------: | :---------: |
43
- | | 5-shot | 5-shot | 5-shot | 0-shot | 3-shot@1 | - | - | - |
44
- | LLaMA2-34B | 62.6 | - | - | - | 44.1 | 69.9 | 68.0 | 26.0 |
45
- | LLaMA2-70B | 68.9 | 53.3 | - | 49.8 | 51.2 | 71.9 | 69.4 | 36.8 |
46
- | Baichuan2-13B | 59.2 | 62.0 | 58.1 | 54.3 | 48.8 | 64.3 | 62.4 | 23.0 |
47
- | Qwen-14B | 66.3 | 71.0 | 72.1 | 62.5 | 53.4 | 73.3 | 72.5 | 39.8 |
48
- | Skywork-13B | 62.1 | 61.8 | 60.6 | 68.1 | 41.7 | 72.4 | 61.4 | 24.9 |
49
- | InternLM-20B | 62.1 | 59.0 | 58.8 | 45.5 | 52.5 | 78.3 | - | 26.0 |
50
- | Aquila-34B | 67.8 | 71.4 | 63.1 | - | - | - | - | - |
51
- | Falcon-180B | 70.4 | 58.0 | 57.8 | 59.0 | 54.0 | 77.3 | 68.8 | 34.0 |
52
- | Yi-6B | 63.2 | 75.5 | 72.0 | 72.2 | 42.8 | 72.3 | 68.7 | 19.8 |
53
- | **Yi-34B** | **76.3** | **83.7** | **81.4** | **82.8** | **54.3** | **80.1** | **76.4** | **37.1** |
54
-
55
-
56
- While benchmarking open-source models, we have observed a disparity between the results generated by our pipeline and those reported in public sources (e.g. OpenCompass). Upon conducting a more in-depth investigation of this difference, we have discovered that various models may employ different prompts, post-processing strategies, and sampling techniques, potentially resulting in significant variations in the outcomes. Our prompt and post-processing strategy remains consistent with the original benchmark, and greedy decoding is employed during evaluation without any post-processing for the generated content. For scores that were not reported by the original authors (including scores reported with different settings), we try to get results with our pipeline.
57
-
58
- To evaluate the model's capability extensively, we adopted the methodology outlined in Llama2. Specifically, we included PIQA, SIQA, HellaSwag, WinoGrande, ARC, OBQA, and CSQA to assess common sense reasoning. SquAD, QuAC, and BoolQ were incorporated to evaluate reading comprehension. CSQA was exclusively tested using a 7-shot setup, while all other tests were conducted with a 0-shot configuration. Additionally, we introduced GSM8K (8-shot@1), MATH (4-shot@1), HumanEval (0-shot@1), and MBPP (3-shot@1) under the category "Math & Code". Due to technical constraints, we did not test Falcon-180 on QuAC and OBQA; the score is derived by averaging the scores on the remaining tasks. Since the scores for these two tasks are generally lower than the average, we believe that Falcon-180B's performance was not underestimated.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
59
 
60
  ## Disclaimer
61
 
62
- Although we use data compliance checking algorithms during the training process to ensure the compliance of the trained model to the best of our ability, due to the complexity of the data and the diversity of language model usage scenarios, we cannot guarantee that the model will generate correct and reasonable output in all scenarios. Please be aware that there is still a risk of the model producing problematic outputs. We will not be responsible for any risks and issues resulting from misuse, misguidance, illegal usage, and related misinformation, as well as any associated data security concerns.
 
 
 
 
 
 
 
63
 
64
  ## License
65
 
66
- The Yi series model must be adhere to the [Model License Agreement](https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE).
67
- For any questions related to licensing and copyright, please contact us ([yi@01.ai](mailto:yi@01.ai)).
 
 
 
 
5
  ---
6
  <div align="center">
7
 
8
+ <img src="./Yi.svg" width="200px">
 
 
9
 
10
  </div>
11
 
12
  ## Introduction
13
 
14
+ The **Yi** series models are large language models trained from scratch by
15
+ developers at [01.AI](https://01.ai/). The first public release contains two
16
+ bilingual(English/Chinese) base models with the parameter sizes of 6B and 34B.
17
+ Both of them are trained with 4K sequence length and can be extended to 32K
18
+ during inference time.
19
 
20
  ## News
21
 
22
+ - 🎯 **2023/11/02**: The base model of `Yi-6B` and `Yi-34B`.
23
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
24
 
25
  ## Model Performance
26
 
27
+ | Model | MMLU | CMMLU | C-Eval | GAOKAO | BBH | Common-sense Reasoning | Reading Comprehension | Math & Code |
28
+ | :------------ | :------: | :------: | :------: | :------: | :------: | :--------------------: | :-------------------: | :---------: |
29
+ | | 5-shot | 5-shot | 5-shot | 0-shot | 3-shot@1 | - | - | - |
30
+ | LLaMA2-34B | 62.6 | - | - | - | 44.1 | 69.9 | 68.0 | 26.0 |
31
+ | LLaMA2-70B | 68.9 | 53.3 | - | 49.8 | 51.2 | 71.9 | 69.4 | 36.8 |
32
+ | Baichuan2-13B | 59.2 | 62.0 | 58.1 | 54.3 | 48.8 | 64.3 | 62.4 | 23.0 |
33
+ | Qwen-14B | 66.3 | 71.0 | 72.1 | 62.5 | 53.4 | 73.3 | 72.5 | **39.8** |
34
+ | Skywork-13B | 62.1 | 61.8 | 60.6 | 68.1 | 41.7 | 72.4 | 61.4 | 24.9 |
35
+ | InternLM-20B | 62.1 | 59.0 | 58.8 | 45.5 | 52.5 | 78.3 | - | 30.4 |
36
+ | Aquila-34B | 67.8 | 71.4 | 63.1 | - | - | - | - | - |
37
+ | Falcon-180B | 70.4 | 58.0 | 57.8 | 59.0 | 54.0 | 77.3 | 68.8 | 34.0 |
38
+ | Yi-6B | 63.2 | 75.5 | 72.0 | 72.2 | 42.8 | 72.3 | 68.7 | 19.8 |
39
+ | **Yi-34B** | **76.3** | **83.7** | **81.4** | **82.8** | **54.3** | **80.1** | **76.4** | 37.1 |
40
+
41
+
42
+ While benchmarking open-source models, we have observed a disparity between the
43
+ results generated by our pipeline and those reported in public sources (e.g.
44
+ OpenCompass). Upon conducting a more in-depth investigation of this difference,
45
+ we have discovered that various models may employ different prompts,
46
+ post-processing strategies, and sampling techniques, potentially resulting in
47
+ significant variations in the outcomes. Our prompt and post-processing strategy
48
+ remains consistent with the original benchmark, and greedy decoding is employed
49
+ during evaluation without any post-processing for the generated content. For
50
+ scores that were not reported by the original authors (including scores reported
51
+ with different settings), we try to get results with our pipeline.
52
+
53
+ To evaluate the model's capability extensively, we adopted the methodology
54
+ outlined in Llama2. Specifically, we included PIQA, SIQA, HellaSwag, WinoGrande,
55
+ ARC, OBQA, and CSQA to assess common sense reasoning. SquAD, QuAC, and BoolQ
56
+ were incorporated to evaluate reading comprehension. CSQA was exclusively tested
57
+ using a 7-shot setup, while all other tests were conducted with a 0-shot
58
+ configuration. Additionally, we introduced GSM8K (8-shot@1), MATH (4-shot@1),
59
+ HumanEval (0-shot@1), and MBPP (3-shot@1) under the category "Math & Code". Due
60
+ to technical constraints, we did not test Falcon-180 on QuAC and OBQA; the score
61
+ is derived by averaging the scores on the remaining tasks. Since the scores for
62
+ these two tasks are generally lower than the average, we believe that
63
+ Falcon-180B's performance was not underestimated.
64
+
65
+ ## Usage
66
+
67
+ Please visit our [github repository](https://github.com/01-ai/Yi) for general
68
+ guidance on how to use this model.
69
 
70
  ## Disclaimer
71
 
72
+ Although we use data compliance checking algorithms during the training process
73
+ to ensure the compliance of the trained model to the best of our ability, due to
74
+ the complexity of the data and the diversity of language model usage scenarios,
75
+ we cannot guarantee that the model will generate correct and reasonable output
76
+ in all scenarios. Please be aware that there is still a risk of the model
77
+ producing problematic outputs. We will not be responsible for any risks and
78
+ issues resulting from misuse, misguidance, illegal usage, and related
79
+ misinformation, as well as any associated data security concerns.
80
 
81
  ## License
82
 
83
+ The **Yi** series models are fully open for academic research and free
84
+ commercial usage. All usage must be adhered to the [Model License
85
+ Agreement](https://huggingface.co/01-ai/Yi-6B/blob/main/LICENSE). To apply for
86
+ the official commercial license, please contact us
87
+ ([yi@01.ai](mailto:yi@01.ai)).
Yi.svg ADDED