🤖 The Yi series models are the next generation of open source large language models trained from strach by 01.AI.
🙌 Targeted as a bilingual language model and trained on 3T multilingual corpus, the Yi series models become one of the strongest LLM worldwide, showing promise in language understanding, commonsense reasoning, reading comprehension, and more. For example,
For English language capability, the Yi series models ranked 2nd (just behind GPT-4), outperforming other LLMs (such as LLaMA2-chat-70B, Claude 2, and ChatGPT) on the AlpacaEval Leaderboard in Dec 2023.
For Chinese language capability, the Yi series models landed in 2nd place (following GPT4), surpassing other LLMs (such as Baidu ERNIE, Qwen, and Baichuan) on the SuperCLUE in Oct 2023.
🙏 (Credits to LLaMA) Thanks to the Transformer and LLaMA open-source communities, as they reducing the efforts required to build from scratch and enabling the utilization of the same tools within the AI ecosystem. If you're interested in Yi's adoption of LLaMA architecture and license usage policy, see Yi's relation with LLaMA.
- 4 bits series models are quantized by AWQ. - 8 bits series models are quantized by GPTQ - All quantized models have a low barrier to use since they can be deployed on consumer-grade GPUs (e.g., 3090, 4090).
- 200k is roughly equivalent to 400,000 Chinese characters.
Other info
For chat models and base models:
6B series models are suitable for personal and academic use.
34B series models suitable for personal, academic, and commercial (particularly for small and medium-sized enterprises) purposes. It's a cost-effective solution that's affordable and equipped with emergent ability.
🎯 2023/11/05: The base model of Yi-6B-200K and Yi-34B-200K.
This release contains two base models with the same parameter sizes of previous
release, except that the context window is extended to 200K.
🎯 2023/11/02: The base model of Yi-6B and Yi-34B.
The first public release contains two bilingual (English/Chinese) base models
with the parameter sizes of 6B and 34B. Both of them are trained with 4K
sequence length and can be extended to 32K during inference time.
The Yi series models follow the same model architecture as LLaMA. By choosing Yi, you can leverage existing tools, libraries, and resources within the LLaMA ecosystem, eliminating the need to create new tools and enhancing development efficiency.
For example, the Yi series models are saved in the format of the LLaMA model. You can directly use LLaMAForCausalLM and LLaMATokenizer to load the model. For more information, see Use the chat model.
Feel free to create a PR and share the fantastic work you've built using the Yi series models.
To help others quickly understand your work, it is recommended to use the format of <model-name>: <model-intro> + <model-highlights>.
🔗 Serving
If you want to get up with Yi in a few minutes, you can use the following services built upon Yi.
Yi-34B-Chat (Yi official beta): you can chat with it. Note that currently it's available through a whitelist. Welcome to apply and experience it firsthand!
Yi-6B-Chat (Replicate): you can use this model with more options by setting additional parameters and calling APIs.
ScaleLLM: you can use this service to run Yi models locally with added flexibility and customization.
⚙️ Quantitation
If you have limited computational capabilities, you can use Yi's quantized models as follows.
These quantized models have reduced precision and but offer increased efficiency, such as faster inference speed and smaller RAM usage.
SUSTech/SUS-Chat-34B: this model ranked first among all models below 70B and outperformed the twice larger deepseek-llm-67b-chat. You can check the result on the Open LLM Leaderboard.
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.
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.
📊 Chat model performance
Model
MMLU
MMLU
CMMLU
CMMLU
C-Eval(val)*
C-Eval(val)*
Truthful QA
BBH
BBH
GSM8k
GSM8k
0-shot
5-shot
0-shot
5-shot
0-shot
5-shot
0-shot
0-shot
3-shot
0-shot
4-shot
LLaMA2-13B-Chat
50.88
47.33
27.47
35.08
27.93
35.88
36.84
32.90
58.22
36.85
2.73
LLaMA2-70B-Chat
59.42
59.86
36.10
40.99
34.99
41.31
53.95
42.36
58.53
47.08
58.68
Baichuan2-13B-Chat
55.09
50.14
58.64
59.47
56.02
54.75
48.98
38.81
47.15
45.72
23.28
Qwen-14B-Chat
63.99
64.98
67.73
70.57
66.12
70.06
52.49
49.65
54.98
59.51
61.18
InternLM-Chat-20B
55.55
57.42
53.55
53.75
51.19
53.57
51.75
42.41
36.68
15.69
43.44
AquilaChat2-34B v1.2
65.15
66.70
67.51
70.02
82.99
89.38
64.33
20.12
34.28
11.52
48.45
Yi-6B-Chat
58.24
60.99
69.44
74.71
68.80
74.22
50.58
39.70
47.15
38.44
44.88
Yi-6B-Chat-8bits(GPTQ)
58.29
60.96
69.21
74.69
69.17
73.85
49.85
40.35
47.26
39.42
44.88
Yi-6B-Chat-4bits(AWQ)
56.78
59.89
67.70
73.29
67.53
72.29
50.29
37.74
43.62
35.71
38.36
Yi-34B-Chat
67.62
73.46
79.11
81.34
77.04
78.53
62.43
51.41
71.74
71.65
75.97
Yi-34B-Chat-8bits(GPTQ)
66.24
73.69
79.05
81.23
76.82
78.97
61.84
52.08
70.97
70.74
75.74
Yi-34B-Chat-4bits(AWQ)
65.77
72.42
78.21
80.50
75.71
77.27
61.84
48.30
69.39
70.51
74.00
We evaluated various benchmarks using both zero-shot and few-shot methods, except for TruthfulQA. Generally, the zero-shot approach is more common in chat models. Our evaluation strategy involves generating responses while following instructions explicitly or implicitly (such as using few-shot examples). We then isolate relevant answers from the generated text. Some models are not well-suited to produce output in the specific format required by instructions in few datasets, which leads to suboptimal results.
*: C-Eval results are evaluated on the validation datasets
📊 Quantized chat model performance
We also provide both 4-bit (AWQ) and 8-bit (GPTQ) quantized Yi chat models. Evaluation results on various benchmarks have shown that the quantized models have negligible losses. Additionally, they reduce the memory footprint size. After testing different configurations of prompts and generation lengths, we highly recommend following the guidelines in the memory footprint table below when selecting a device to run our models.
batch=1
batch=4
batch=16
batch=32
Yi-34B-Chat
65GiB
68GiB
76GiB
>80GiB
Yi-34B-Chat-8bits(GPTQ)
35GiB
37GiB
46GiB
58GiB
Yi-34B-Chat-4bits(AWQ)
19GiB
20GiB
30GiB
40GiB
Yi-6B-Chat
12GiB
13GiB
15GiB
18GiB
Yi-6B-Chat-8bits(GPTQ)
7GiB
8GiB
10GiB
14GiB
Yi-6B-Chat-4bits(AWQ)
4GiB
5GiB
7GiB
10GiB
Note: All the numbers in the table represent the minimum recommended memory for running models of the corresponding size.
⛔️ Limitations of chat model
The released chat model has undergone exclusive training using Supervised Fine-Tuning (SFT). Compared to other standard chat models, our model produces more diverse responses, making it suitable for various downstream tasks, such as creative scenarios. Furthermore, this diversity is expected to enhance the likelihood of generating higher quality responses, which will be advantageous for subsequent Reinforcement Learning (RL) training.
However, this higher diversity might amplify certain existing issues, including:
Hallucination: This refers to the model generating factually incorrect or nonsensical information. With the model's responses being more varied, there's a higher chance of hallucination that are not based on accurate data or logical reasoning.
Non-determinism in re-generation: When attempting to regenerate or sample responses, inconsistencies in the outcomes may occur. The increased diversity can lead to varying results even under similar input conditions.
Cumulative Error: This occurs when errors in the model's responses compound over time. As the model generates more diverse responses, the likelihood of small inaccuracies building up into larger errors increases, especially in complex tasks like extended reasoning, mathematical problem-solving, etc.
To achieve more coherent and consistent responses, it is advisable to adjust generation configuration parameters such astemperature,top_p, ortop_k. These adjustments can help in the balance between creativity and coherence in the model's outputs.
The best approach to try the Yi series models is through Docker with GPUs. We
provide the following docker images to help you get started.
registry.lingyiwanwu.com/ci/01-ai/yi:latest
ghcr.io/01-ai/yi:latest
Note that the latest tag always points to the latest code in the main
branch. To test a stable version, please replace it with a specific
tag.
1.2 Local development environment
We use conda-lock to generate fully reproducible lock files for conda environments. You can refer to conda-lock.yml for the exact versions of the dependencies. Additionally, we utilize micromamba for installing these dependencies.
To install the dependencies, please follow these steps:
Install micromamba by following the instructions available here.
Execute micromamba install -y -n yi -f conda-lock.yml to create a conda environment named yi and install the necessary dependencies.
2. Download the model (optional)
By default, the model weights and tokenizer will be downloaded from
Hugging Face automatically in the next step. You
can also download them manually from the following places:
1from transformers import AutoModelForCausalLM, AutoTokenizer
23model_path ='01-ai/Yi-34b-Chat'45tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)67# Since transformers 4.35.0, the GPT-Q/AWQ model can be loaded using AutoModelForCausalLM.8model = AutoModelForCausalLM.from_pretrained(9 model_path,10 device_map="auto",11 torch_dtype='auto'12).eval()1314# Prompt content: "hi"15messages =[16{"role":"user","content":"hi"}17]1819input_ids = tokenizer.apply_chat_template(conversation=messages, tokenize=True, add_generation_prompt=True, return_tensors='pt')20output_ids = model.generate(input_ids.to('cuda'))21response = tokenizer.decode(output_ids[0][input_ids.shape[1]:], skip_special_tokens=True)2223# Model response: "Hello! How can I assist you today?"24print(response)
To construct the prompt template manually, you can refer the chat_template field in the tokenizer_config.jsonfile.
1from transformers import AutoModelForCausalLM, AutoTokenizer
23model = AutoModelForCausalLM.from_pretrained("01-ai/Yi-34B", device_map="auto", torch_dtype="auto")4tokenizer = AutoTokenizer.from_pretrained("01-ai/Yi-34B")5inputs = tokenizer("There's a place where time stands still. A place of breath taking wonder, but also", return_tensors="pt")6max_length =25678outputs = model.generate(9 inputs.input_ids.cuda(),10 max_length=max_length,11 eos_token_id=tokenizer.eos_token_id,12 do_sample=True,13 repetition_penalty=1.3,14 no_repeat_ngram_size=5,15 temperature=0.7,16 top_k=40,17 top_p=0.8,18)19print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Output
Prompt: There's a place where time stands still. A place of breath taking wonder, but also
Generation: There's a place where time stands still. A place of breath taking wonder, but also of great danger. A place where the very air you breathe could kill you. A place where the only way to survive is to be prepared.
The place is called the Arctic.
The Arctic is a vast, frozen wilderness. It is a place of extremes. The temperatures can drop to -40 degrees Celsius. The winds can reach speeds of 100 kilometers per hour. The sun can shine for 24 hours a day, or not at all for weeks on end.
The Arctic is also a place of great beauty. The ice and snow are a pristine white. The sky is a deep blue. The sunsets are spectacular.
But the Arctic is also a place of great danger. The ice can be treacherous. The winds can be deadly. The sun can be blinding.
The Arctic is a place where the only way to survive is to be prepared.
The Arctic is a place of extremes. The temperatures can drop to -40 degrees Celsius. The winds can reach speeds of 100 kilometers per hour. The sun can shine for 24 hours a day, or not at all for weeks on end.
The Arctic is a place of great beauty. The ice and snow are a
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
complex 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.
