These quantised GGUFv2 files are compatible with llama.cpp from August 27th onwards, as of commit d0cee0d
They are also compatible with many third party UIs and libraries - please see the list at the top of this README.
Explanation of quantisation methods
Click to see details
The new methods available are:
GGML_TYPE_Q2_K - "type-1" 2-bit quantization in super-blocks containing 16 blocks, each block having 16 weight. Block scales and mins are quantized with 4 bits. This ends up effectively using 2.5625 bits per weight (bpw)
GGML_TYPE_Q3_K - "type-0" 3-bit quantization in super-blocks containing 16 blocks, each block having 16 weights. Scales are quantized with 6 bits. This end up using 3.4375 bpw.
GGML_TYPE_Q4_K - "type-1" 4-bit quantization in super-blocks containing 8 blocks, each block having 32 weights. Scales and mins are quantized with 6 bits. This ends up using 4.5 bpw.
GGML_TYPE_Q5_K - "type-1" 5-bit quantization. Same super-block structure as GGML_TYPE_Q4_K resulting in 5.5 bpw
GGML_TYPE_Q6_K - "type-0" 6-bit quantization. Super-blocks with 16 blocks, each block having 16 weights. Scales are quantized with 8 bits. This ends up using 6.5625 bpw
Refer to the Provided Files table below to see what files use which methods, and how.
very large, extremely low quality loss - not recommended
Note: the above RAM figures assume no GPU offloading. If layers are offloaded to the GPU, this will reduce RAM usage and use VRAM instead.
How to download GGUF files
Note for manual downloaders: You almost never want to clone the entire repo! Multiple different quantisation formats are provided, and most users only want to pick and download a single file.
The following clients/libraries will automatically download models for you, providing a list of available models to choose from:
LM Studio
LoLLMS Web UI
Faraday.dev
In text-generation-webui
Under Download Model, you can enter the model repo: TheBloke/Yi-34B-Chat-GGUF and below it, a specific filename to download, such as: yi-34b-chat.Q4_K_M.gguf.
Then click Download.
On the command line, including multiple files at once
I recommend using the huggingface-hub Python library:
pip3 install huggingface-hub
Then you can download any individual model file to the current directory, at high speed, with a command like this:
Change -ngl 32 to the number of layers to offload to GPU. Remove it if you don't have GPU acceleration.
Change -c 2048 to the desired sequence length. For extended sequence models - eg 8K, 16K, 32K - the necessary RoPE scaling parameters are read from the GGUF file and set by llama.cpp automatically.
If you want to have a chat-style conversation, replace the -p <PROMPT> argument with -i -ins
How to load this model in Python code, using ctransformers
First install the package
Run one of the following commands, according to your system:
shell
1# Base ctransformers with no GPU acceleration2pip install ctransformers
3# Or with CUDA GPU acceleration4pip install ctransformers[cuda]5# Or with AMD ROCm GPU acceleration (Linux only)6CT_HIPBLAS=1 pip install ctransformers --no-binary ctransformers
7# Or with Metal GPU acceleration for macOS systems only8CT_METAL=1 pip install ctransformers --no-binary ctransformers
Simple ctransformers example code
python
1from ctransformers import AutoModelForCausalLM
23# Set gpu_layers to the number of layers to offload to GPU. Set to 0 if no GPU acceleration is available on your system.4llm = AutoModelForCausalLM.from_pretrained("TheBloke/Yi-34B-Chat-GGUF", model_file="yi-34b-chat.Q4_K_M.gguf", model_type="yi", gpu_layers=50)56print(llm("AI is going to"))
How to use with LangChain
Here are guides on using llama-cpp-python and ctransformers with LangChain:
I've had a lot of people ask if they can contribute. I enjoy providing models and helping people, and would love to be able to spend even more time doing it, as well as expanding into new projects like fine tuning/training.
If you're able and willing to contribute it will be most gratefully received and will help me to keep providing more models, and to start work on new AI projects.
Donaters will get priority support on any and all AI/LLM/model questions and requests, access to a private Discord room, plus other benefits.
Patreon special mentions: Brandon Frisco, LangChain4j, Spiking Neurons AB, transmissions 11, Joseph William Delisle, Nitin Borwankar, Willem Michiel, Michael Dempsey, vamX, Jeffrey Morgan, zynix, jjj, Omer Bin Jawed, Sean Connelly, jinyuan sun, Jeromy Smith, Shadi, Pawan Osman, Chadd, Elijah Stavena, Illia Dulskyi, Sebastain Graf, Stephen Murray, terasurfer, Edmond Seymore, Celu Ramasamy, Mandus, Alex, biorpg, Ajan Kanaga, Clay Pascal, Raven Klaugh, 阿明, K, ya boyyy, usrbinkat, Alicia Loh, John Villwock, ReadyPlayerEmma, Chris Smitley, Cap'n Zoog, fincy, GodLy, S_X, sidney chen, Cory Kujawski, OG, Mano Prime, AzureBlack, Pieter, Kalila, Spencer Kim, Tom X Nguyen, Stanislav Ovsiannikov, Michael Levine, Andrey, Trailburnt, Vadim, Enrico Ros, Talal Aujan, Brandon Phillips, Jack West, Eugene Pentland, Michael Davis, Will Dee, webtim, Jonathan Leane, Alps Aficionado, Rooh Singh, Tiffany J. Kim, theTransient, Luke @flexchar, Elle, Caitlyn Gatomon, Ari Malik, subjectnull, Johann-Peter Hartmann, Trenton Dambrowitz, Imad Khwaja, Asp the Wyvern, Emad Mostaque, Rainer Wilmers, Alexandros Triantafyllidis, Nicholas, Pedro Madruga, SuperWojo, Harry Royden McLaughlin, James Bentley, Olakabola, David Ziegler, Ai Maven, Jeff Scroggin, Nikolai Manek, Deo Leter, Matthew Berman, Fen Risland, Ken Nordquist, Manuel Alberto Morcote, Luke Pendergrass, TL, Fred von Graf, Randy H, Dan Guido, NimbleBox.ai, Vitor Caleffi, Gabriel Tamborski, knownsqashed, Lone Striker, Erik Bjäreholt, John Detwiler, Leonard Tan, Iucharbius
Thank you to all my generous patrons and donaters!
And thank you again to a16z for their generous grant.
Original model card: 01-ai's Yi 34B Chat
Introduction
The Yi series models are large language models trained from scratch by
developers at 01.AI.
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 quntinized by GPTQ, two 4-bits models quantinized by AWQ.
🎯 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.
Model Performance
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.
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.
Usage
Feel free to create an issue if you
encounter any problem when using the Yi series models.
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:
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
HuggingFace 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)
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:
1from transformers import AutoModelForCausalLM, AutoTokenizer
23model = AutoModelForCausalLM.from_pretrained("01-ai/Yi-34B", device_map="auto", torch_dtype="auto", trust_remote_code=True)4tokenizer = AutoTokenizer.from_pretrained("01-ai/Yi-34B", trust_remote_code=True)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
The dataset we use contains Chinese & English only. We used approximately 3T
tokens. The detailed number and its construction will be described in the
upcoming technical report.
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.
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 Model License
Agreement 2.0.
To apply for the official commercial license, please contact us
(yi@01.ai).
