Edit model card

Model description

This model is a fine-tuned version of 01-ai/Yi-6B-Chat on the alpaca_gpt4_en dataset.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric Value
Avg. 59.83
AI2 Reasoning Challenge (25-Shot) 56.57
HellaSwag (10-Shot) 77.66
MMLU (5-Shot) 64.05
TruthfulQA (0-shot) 50.75
Winogrande (5-shot) 73.95
GSM8k (5-shot) 36.01

ORIGINAL MODEL CARD:


Building the Next Generation of Open-Source and Bilingual LLMs

πŸ€— Hugging Face β€’ πŸ€– ModelScope β€’ ✑️ WiseModel

πŸ‘‹ Join us πŸ’¬ WeChat (Chinese) !


πŸ“• Table of Contents

🟒 What is Yi?

πŸ“Œ Introduction

  • πŸ€– The Yi series models are the next generation of open source large language models trained from strach by 01.AI.

  • πŸ™Œ 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.

  • πŸ™ (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.

🎯 Models

Yi models come in multiple sizes and cater to different use cases. You can also fine-tune Yi models to meet your specific requirements.

Chat models

Model Download
Yi-6B-Chat β€’ πŸ€— Hugging Face β€’ πŸ€– ModelScope
Yi-6B-Chat-4bits β€’ πŸ€— Hugging Face β€’ πŸ€– ModelScope
Yi-6B-Chat-8bits β€’ πŸ€— Hugging Face β€’ πŸ€– ModelScope
Yi-34B-Chat β€’ πŸ€— Hugging Face β€’ πŸ€– ModelScope
Yi-34B-Chat-4bits β€’ πŸ€— Hugging Face β€’ πŸ€– ModelScope
Yi-34B-Chat-8bits β€’ πŸ€— Hugging Face β€’ πŸ€– ModelScope

- 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).

Base models

Model Download
Yi-6B β€’ πŸ€— Hugging Face β€’ πŸ€– ModelScope
Yi-6B-200K β€’ πŸ€— Hugging Face β€’ πŸ€– ModelScope
Yi-34B β€’ πŸ€— Hugging Face β€’ πŸ€– ModelScope
Yi-34B-200K β€’ πŸ€— Hugging Face β€’ πŸ€– ModelScope

- 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.

  • The default context window is 4k tokens.

  • The pretrained tokens are 3T.

  • The training data are up to June 2023.

πŸŽ‰ News

🎯 2023/11/23: The chat models are open to public.

This release contains two chat models based on previous released base models, two 8-bits models quantized by GPTQ, two 4-bits models quantized by AWQ.

  • Yi-34B-Chat
  • Yi-34B-Chat-4bits
  • Yi-34B-Chat-8bits
  • Yi-6B-Chat
  • Yi-6B-Chat-4bits
  • Yi-6B-Chat-8bits

You can try some of them interactively at:

πŸ”” 2023/11/23: The Yi Series Models Community License Agreement is updated to v2.1.
πŸ”₯ 2023/11/08: Invited test of Yi-34B chat model.

Application form:

🎯 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.

🟒 Why Yi?

🌎 Ecosystem

Yi has a comprehensive ecosystem, offering a range of tools, services, and models to enrich your experiences and maximize productivity.

πŸ’¦ Upstream

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.

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("01-ai/Yi-34b", use_fast=False)

model = AutoModelForCausalLM.from_pretrained("01-ai/Yi-34b", device_map="auto")

🌊 Downstream

πŸ’‘ Tip

  • 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>.

πŸ”— 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.

βš™οΈ 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.

πŸ› οΈ Fine-tuning

If you're seeking to explore the diverse capabilities within Yi's thriving family, you can delve into Yi's fine-tuned models as below.

πŸ“Œ Benchmarks

πŸ“Š Base model performance

Model MMLU CMMLU C-Eval GAOKAO BBH Common-sense Reasoning Reading Comprehension Math & Code
5-shot 5-shot 5-shot 0-shot 3-shot@1 - - -
LLaMA2-34B 62.6 - - - 44.1 69.9 68.0 26.0
LLaMA2-70B 68.9 53.3 - 49.8 51.2 71.9 69.4 36.8
Baichuan2-13B 59.2 62.0 58.1 54.3 48.8 64.3 62.4 23.0
Qwen-14B 66.3 71.0 72.1 62.5 53.4 73.3 72.5 39.8
Skywork-13B 62.1 61.8 60.6 68.1 41.7 72.4 61.4 24.9
InternLM-20B 62.1 59.0 58.8 45.5 52.5 78.3 - 30.4
Aquila-34B 67.8 71.4 63.1 - - - - -
Falcon-180B 70.4 58.0 57.8 59.0 54.0 77.3 68.8 34.0
Yi-6B 63.2 75.5 72.0 72.2 42.8 72.3 68.7 19.8
Yi-6B-200K 64.0 75.3 73.5 73.9 42.0 72.0 69.1 19.0
Yi-34B 76.3 83.7 81.4 82.8 54.3 80.1 76.4 37.1
Yi-34B-200K 76.1 83.6 81.9 83.4 52.7 79.7 76.6 36.3

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.

πŸ“Š 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

πŸ“Š 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.

🟒 Who can use Yi?

Everyone! πŸ™Œ βœ…

🟒 How to use Yi?

1. Prepare development environment
2. Download the model
3. Examples

1. Prepare development environment

1.1 Docker

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:

  1. Install micromamba by following the instructions available here.
  2. 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:

3. Examples

3.1 Use the chat model

from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = '01-ai/Yi-34b-Chat'

tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)

# Since transformers 4.35.0, the GPT-Q/AWQ model can be loaded using AutoModelForCausalLM.
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    device_map="auto",
    torch_dtype='auto'
).eval()

# Prompt content: "hi"
messages = [
    {"role": "user", "content": "hi"}
]

input_ids = tokenizer.apply_chat_template(conversation=messages, tokenize=True, add_generation_prompt=True, return_tensors='pt')
output_ids = model.generate(input_ids.to('cuda'))
response = tokenizer.decode(output_ids[0][input_ids.shape[1]:], skip_special_tokens=True)

# Model response: "Hello! How can I assist you today?"
print(response)

To construct the prompt template manually, you can refer the chat_template field in the tokenizer_config.json file.

<|im_start|>system
{system_message}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

3.2 Use the base model

python demo/text_generation.py

To reuse the downloaded models in the previous step, you can provide the extra --model argument:

python demo/text_generation.py  --model /path/to/model

Or if you'd like to get your hands dirty:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("01-ai/Yi-34B", device_map="auto", torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained("01-ai/Yi-34B")
inputs = tokenizer("There's a place where time stands still. A place of breath taking wonder, but also", return_tensors="pt")
max_length = 256

outputs = model.generate(
    inputs.input_ids.cuda(),
    max_length=max_length,
    eos_token_id=tokenizer.eos_token_id,
    do_sample=True,
    repetition_penalty=1.3,
    no_repeat_ngram_size=5,
    temperature=0.7,
    top_k=40,
    top_p=0.8,
)
print(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

For more advanced usage, please refer to the doc.

3.3 Finetune from the base model

bash finetune/scripts/run_sft_Yi_6b.sh

Once finished, you can compare the finetuned model and the base model with the following command:

bash finetune/scripts/run_eval.sh

For more advanced usage like fine-tuning based on your custom data, please refer the doc.

3.4 Quantization

GPT-Q
python quantization/gptq/quant_autogptq.py \
  --model /base_model                      \
  --output_dir /quantized_model            \
  --trust_remote_code

Once finished, you can then evaluate the resulting model as follows:

python quantization/gptq/eval_quantized_model.py \
  --model /quantized_model                       \
  --trust_remote_code

For a more detailed explanation, please read the doc

AWQ
python quantization/awq/quant_autoawq.py \
  --model /base_model                      \
  --output_dir /quantized_model            \
  --trust_remote_code

Once finished, you can then evaluate the resulting model as follows:

python quantization/awq/eval_quantized_model.py \
  --model /quantized_model                       \
  --trust_remote_code

For more detailed explanation, please read the doc

🟒 Misc.

πŸ“‘ Disclaimer

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.

πŸͺͺ License

The source code in this repo is licensed under the Apache 2.0 license. The Yi series models are fully open for academic research and free commercial usage with permission via applications. All usage must adhere to the Yi Series Models Community License Agreement 2.1. For free commercial use, you only need to send an email to get official commercial permission.

Downloads last month
4,223
Safetensors
Model size
6.06B params
Tensor type
F32
Β·
FP16
Β·
Inference Examples
Inference API (serverless) has been turned off for this model.

Model tree for Walmart-the-bag/Yi-6B-Infinity-Chat

Base model

01-ai/Yi-6B-Chat
Finetuned
(11)
this model
Quantizations
1 model

Collection including Walmart-the-bag/Yi-6B-Infinity-Chat

Evaluation results