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: Sam, theTransient, Jonathan Leane, Steven Wood, webtim, Johann-Peter Hartmann, Geoffrey Montalvo, Gabriel Tamborski, Willem Michiel, John Villwock, Derek Yates, Mesiah Bishop, Eugene Pentland, Pieter, Chadd, Stephen Murray, Daniel P. Andersen, terasurfer, Brandon Frisco, Thomas Belote, Sid, Nathan LeClaire, Magnesian, Alps Aficionado, Stanislav Ovsiannikov, Alex, Joseph William Delisle, Nikolai Manek, Michael Davis, Junyu Yang, K, J, Spencer Kim, Stefan Sabev, Olusegun Samson, transmissions 11, Michael Levine, Cory Kujawski, Rainer Wilmers, zynix, Kalila, Luke @flexchar, Ajan Kanaga, Mandus, vamX, Ai Maven, Mano Prime, Matthew Berman, subjectnull, Vitor Caleffi, Clay Pascal, biorpg, alfie_i, 阿明, Jeffrey Morgan, ya boyyy, Raymond Fosdick, knownsqashed, Olakabola, Leonard Tan, ReadyPlayerEmma, Enrico Ros, Dave, Talal Aujan, Illia Dulskyi, Sean Connelly, senxiiz, Artur Olbinski, Elle, Raven Klaugh, Fen Risland, Deep Realms, Imad Khwaja, Fred von Graf, Will Dee, usrbinkat, SuperWojo, Alexandros Triantafyllidis, Swaroop Kallakuri, Dan Guido, John Detwiler, Pedro Madruga, Iucharbius, Viktor Bowallius, Asp the Wyvern, Edmond Seymore, Trenton Dambrowitz, Space Cruiser, Spiking Neurons AB, Pyrater, LangChain4j, Tony Hughes, Kacper Wikieł, Rishabh Srivastava, David Ziegler, Luke Pendergrass, Andrey, Gabriel Puliatti, Lone Striker, Sebastain Graf, Pierre Kircher, Randy H, NimbleBox.ai, Vadim, danny, Deo Leter
Thank you to all my generous patrons and donaters!
And thank you again to a16z for their generous grant.
The Guanaco models are open-source finetuned chatbots obtained through 4-bit QLoRA tuning of LLaMA base models on the OASST1 dataset. They are available in 7B, 13B, 33B, and 65B parameter sizes.
⚠️Guanaco is a model purely intended for research purposes and could produce problematic outputs.
Why use Guanaco?
Competitive with commercial chatbot systems on the Vicuna and OpenAssistant benchmarks (ChatGPT and BARD) according to human and GPT-4 raters. We note that the relative performance on tasks not covered in these benchmarks could be very different. In addition, commercial systems evolve over time (we used outputs from the March 2023 version of the models).
Available open-source for research purposes. Guanaco models allow cheap and local experimentation with high-quality chatbot systems.
Replicable and efficient training procedure that can be extended to new use cases. Guanaco training scripts are available in the QLoRA repo.
Rigorous comparison to 16-bit methods (both 16-bit full-finetuning and LoRA) in our paper demonstrates the effectiveness of 4-bit QLoRA finetuning.
Lightweight checkpoints which only contain adapter weights.
License and Intended Use
Guanaco adapter weights are available under Apache 2 license. Note the use of the Guanaco adapter weights, requires access to the LLaMA model weighs.
Guanaco is based on LLaMA and therefore should be used according to the LLaMA license.
Usage
Here is an example of how you would load Guanaco 7B in 4-bits:
Inference can then be performed as usual with HF models as follows:
python
1prompt ="Introduce yourself"2formatted_prompt =(3f"A chat between a curious human and an artificial intelligence assistant."4f"The assistant gives helpful, detailed, and polite answers to the user's questions.\n"5f"### Human: {prompt} ### Assistant:"6)7inputs = tokenizer(formatted_prompt, return_tensors="pt").to("cuda:0")8outputs = model.generate(inputs=inputs.input_ids, max_new_tokens=20)9print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Expected output similar to the following:
A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.
### Human: Introduce yourself ### Assistant: I am an artificial intelligence assistant. I am here to help you with any questions you may have.
Current Inference Limitations
Currently, 4-bit inference is slow. We recommend loading in 16 bits if inference speed is a concern. We are actively working on releasing efficient 4-bit inference kernels.
Architecture: The Guanaco models are LoRA adapters to be used on top of LLaMA models. They are added to all layers. For all model sizes, we use $r=64$.
Base Model: Guanaco uses LLaMA as base model with sizes 7B, 13B, 33B, 65B. LLaMA is a causal language model pretrained on a large corpus of text. See LLaMA paper for more details. Note that Guanaco can inherit biases and limitations of the base model.
Languages: The OASST1 dataset is multilingual (see the paper for details) and as such Guanaco responds to user queries in different languages. We note, however, that OASST1 is heavy in high-resource languages. In addition, human evaluation of Guanaco was only performed in English and based on qualitative analysis we observed degradation in performance in other languages.
Next, we describe Training and Evaluation details.
Training
Guanaco models are the result of 4-bit QLoRA supervised finetuning on the OASST1 dataset.
All models use NormalFloat4 datatype for the base model and LoRA adapters on all linear layers with BFloat16 as computation datatype. We set LoRA $r=64$, $\alpha=16$. We also use Adam beta2 of 0.999, max grad norm of 0.3 and LoRA dropout of 0.1 for models up to 13B and 0.05 for 33B and 65B models.
For the finetuning process, we use constant learning rate schedule and paged AdamW optimizer.
Training hyperparameters
Size
Dataset
Batch Size
Learning Rate
Max Steps
Sequence length
7B
OASST1
16
2e-4
1875
512
13B
OASST1
16
2e-4
1875
512
33B
OASST1
16
1e-4
1875
512
65B
OASST1
16
1e-4
1875
512
Evaluation
We test generative language capabilities through both automated and human evaluations. This second set of evaluations relies on queries curated by humans and aims at measuring the quality of model responses. We use the Vicuna and OpenAssistant datasets with 80 and 953 prompts respectively.
In both human and automated evaluations, for each prompt, raters compare all pairs of responses across the models considered. For human raters we randomize the order of the systems, for GPT-4 we evaluate with both orders.
Benchmark
Vicuna
Vicuna
OpenAssistant
-
Prompts
80
80
953
Judge
Human
GPT-4
GPT-4
Model
Elo
Rank
Elo
Rank
Elo
Rank
Median Rank
GPT-4
1176
1
1348
1
1294
1
1
Guanaco-65B
1023
2
1022
2
1008
3
2
Guanaco-33B
1009
4
992
3
1002
4
4
ChatGPT-3.5 Turbo
916
7
966
5
1015
2
5
Vicuna-13B
984
5
974
4
936
5
5
Guanaco-13B
975
6
913
6
885
6
6
Guanaco-7B
1010
3
879
8
860
7
7
Bard
909
8
902
7
-
-
8
We also use the MMLU benchmark to measure performance on a range of language understanding tasks. This is a multiple-choice benchmark covering 57 tasks including elementary mathematics, US history, computer science, law, and more. We report 5-shot test accuracy.
Dataset
7B
13B
33B
65B
LLaMA no tuning
35.1
46.9
57.8
63.4
Self-Instruct
36.4
33.3
53.0
56.7
Longform
32.1
43.2
56.6
59.7
Chip2
34.5
41.6
53.6
59.8
HH-RLHF
34.9
44.6
55.8
60.1
Unnatural Instruct
41.9
48.1
57.3
61.3
OASST1 (Guanaco)
36.6
46.4
57.0
62.2
Alpaca
38.8
47.8
57.3
62.5
FLAN v2
44.5
51.4
59.2
63.9
Risks and Biases
The model can produce factually incorrect output, and should not be relied on to produce factually accurate information. The model was trained on various public datasets; it is possible that this model could generate lewd, biased, or otherwise offensive outputs.
However, we note that finetuning on OASST1 seems to reduce biases as measured on the CrowS dataset. We report here the performance of Guanaco-65B compared to other baseline models on the CrowS dataset.
LLaMA-65B
GPT-3
OPT-175B
Guanaco-65B
Gender
70.6
62.6
65.7
47.5
Religion
{79.0}
73.3
68.6
38.7
Race/Color
57.0
64.7
68.6
45.3
Sexual orientation
{81.0}
76.2
78.6
59.1
Age
70.1
64.4
67.8
36.3
Nationality
64.2
61.6
62.9
32.4
Disability
66.7
76.7
76.7
33.9
Physical appearance
77.8
74.6
76.2
43.1
Socioeconomic status
71.5
73.8
76.2
55.3
Average
66.6
67.2
69.5
43.5
Citation
bibtex
1@article{dettmers2023qlora,
2 title={QLoRA: Efficient Finetuning of Quantized LLMs},
3 author={Dettmers, Tim and Pagnoni, Artidoro and Holtzman, Ari and Zettlemoyer, Luke},
4 journal={arXiv preprint arXiv:2305.14314},
5 year={2023}
6}
