Skip to content

Stability-AI/StableLM

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

32 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

StableLM: Stability AI Language Models

Stochastic Parrot
“A Stochastic Parrot, flat design, vector art” — Stable Diffusion XL

This repository contains Stability AI's ongoing development of the StableLM series of language models and will be continuously updated with new checkpoints. The following provides an overview of all currently available models. More coming soon.

News

September 29, 2023

August 5, 2023

  • Released patched StableLM-Alpha v2 models with 3B and 7B parameters.

April 28, 2023

  • Released StableVicuna-13B, our RLHF fine-tune of Vicuna-13B v0, which itself is a fine-tune of LLaMA-13B. Delta weights over the original Llama model is released under (CC BY-NC-SA-4.0).

April 20, 2023

  • Released initial set of StableLM-Alpha models, with 3B and 7B parameters. Base models are released under CC BY-SA-4.0.

  • Try to chat with our 7B model, StableLM-Tuned-Alpha-7B, on Hugging Face Spaces.

Models

StableLM-3B-4E1T

Technical Report: StableLM-3B-4E1T

StableLM-3B-4E1T is a 3 billion (3B) parameter language model pre-trained under the multi-epoch regime to study the impact of repeated tokens on downstream performance. Given prior success in this area (Tay et al., 2023 and Taylor et al., 2022), we train on 1 trillion (1T) tokens for 4 epochs following the observations of Muennighoff et al. (2023) in "Scaling Data-Constrained Language Models" in which they find "training with up to 4 epochs of repeated data yields negligible changes to loss compared to having unique data." Further inspiration for the token count is taken from "Go smol or go home" (De Vries, 2023), which suggests a 2.96B model trained for 2.85 trillion tokens achieves a similar loss to a Chinchilla compute-optimal 9.87B language model ($k_n = 0.3$).

Size StableLM-3B-4E1T Training Tokens Parameters
3B checkpoint 4T 2,795,443,200

Model Architecture

The model is a decoder-only transformer similar to the LLaMA (Touvron et al., 2023) architecture with the following modifications:

Parameters Hidden Size Layers Heads Sequence Length
2,795,443,200 2560 32 32 4096

Training Data

The dataset is comprised of a filtered mixture of open-source large-scale datasets available on the HuggingFace Hub: Falcon RefinedWeb extract (Penedo et al., 2023), and RedPajama-Data (Together Computer., 2023) and The Pile (Gao et al., 2020) both without Books3 and other subsets, and StarCoder (Li et al., 2023).

Given the large amount of web data, we recommend fine-tuning the base StableLM-3B-4E1T for your downstream tasks.

Training Details

Please refer to the provided YAML configuration file stablelm-3b-4e1t.yml for complete hyperparameter settings and the technical report for further details.

Downstream Results

The following zero-shot evaluations are performed with the lm-evaluation-harness using the lm-bench branch of Stability AI's fork. Full lm-eval JSONs can be found in the evals directory.

Pre-Trained Model Average ARC
Challenge
ARC
Easy
BoolQ HellaSwag (✱) LAMBADA
OpenAI
OpenBookQA PIQA SciQ Winogrande
meta-llama/Llama-2-13b-hf 71.77 48.63 79.50 80.52 79.36 76.77 35.40 79.05 94.50 72.22
huggyllama/llama-7b 68.84 41.89 75.25 75.05 76.22 73.55 34.40 78.67 94.60 69.93
meta-llama/Llama-2-7b-hf 68.75 43.00 76.26 77.74 75.94 73.47 31.40 77.75 93.60 69.61
Qwen/Qwen-7B 67.91 45.39 67.38 74.56 88.85 (?) 69.67 32.20 73.99 93.20 65.98
tiiuae/falcon-7b 67.83 40.27 74.41 73.55 76.35 74.56 30.60 79.49 94.00 67.25
mosaicml/mpt-7b 67.36 40.53 74.92 73.94 76.17 68.64 31.40 78.89 93.70 68.03
stabilityai/stablelm-3b-4e1t 66.93 37.80 72.47 75.63 73.90 70.64 31.40 79.22 94.80 66.54
baichuan-inc/Baichuan2-7B-Base 66.93 42.24 75.00 73.09 72.29 70.99 30.40 76.17 94.60 67.56
stabilityai/stablelm-base-alpha-7b-v2 66.89 38.48 73.19 70.31 74.27 74.19 30.40 78.45 93.90 68.82
openlm-research/open_llama_7b_v2 66.32 38.82 71.93 71.41 74.65 71.05 30.20 79.16 93.80 65.82
microsoft/phi-1_5 65.57 44.45 76.14 74.53 62.62 52.75 37.60 76.33 93.20 72.53
EleutherAI/gpt-neox-20B 65.57 37.88 72.90 69.48 71.43 71.98 29.80 77.42 93.10 66.14
togethercomputer/RedPajama-INCITE-7B-Base 65.07 37.71 72.35 70.76 70.33 71.34 29.00 77.15 92.70 64.33
cerebras/btlm-3b-8k-base (§) 63.59 34.90 70.45 69.63 69.78 66.23 27.60 75.84 92.90 64.96
EleutherAI/pythia-12b 62.69 31.83 70.20 67.31 67.38 70.64 26.40 76.28 90.20 64.01
openlm-research/open_llama_3b_v2 62.43 33.87 67.59 65.69 69.99 66.74 26.00 76.66 92.40 62.90
EleutherAI/gpt-j-6B 62.34 33.96 66.96 65.44 66.24 68.23 29.00 75.57 91.50 64.17
stabilityai/stablelm-base-alpha-3b-v2 62.19 32.42 67.26 64.56 68.58 70.25 26.40 76.01 92.10 62.12
facebook/opt-6.7b 61.85 30.72 65.66 66.02 67.20 67.65 27.60 76.33 90.10 65.35
EleutherAI/pythia-6.9b 60.58 31.83 67.21 64.01 63.88 67.01 25.80 75.08 89.80 60.62
EleutherAI/pythia-2.8b-deduped 58.52 30.12 63.47 64.13 59.44 65.15 23.80 74.10 88.20 58.25
§ Previous 3B Pre-Trained SOTA
? Outlier Reuslts
* Byte-length Normalized Accuracy

StableLM-3B-4E1T achieves state-of-the-art performance (September 2023) at the 3B parameter scale for open-source models and is competitive with many of the popular contemporary 7B models, even outperforming our most recent 7B StableLM-Base-Alpha-v2.

StableLM-Alpha v2

StableLM-Alpha v2 models significantly improve on the initial Alpha models by incorporating architectural improvements such as SwiGLU (Shazeer, 2020) and using higher-quality data sources, as discussed below. The context length for these models is 4096 tokens.

Size StableLM-Base-Alpha-v2 Training Tokens Parameters
3B checkpoint 1.1T 2,796,431,360
7B checkpoint 1.1T 6,890,209,280

Training Details

Please refer to the provided YAML configuration files for hyperparameter details. E.g. for the extended StableLM-Alpha-3B-v2 model, see stablelm-base-alpha-3b-v2-4k-extension.yml.

Following similar work, we use a multi-stage approach to context length extension (Nijkamp et al., 2023), scheduling 1 trillion tokens at context length 2048 followed by 100 billion tokens at 4096. We found that sequence length warmup (Li et al., 2022) helped stabilize early spikes during the first ~80 billion tokens of pre-training. However, it was not applied to the final runs due to significant throughput penalties as length shapes grew across the curriculum.

Training Data

The most impactful changes for StableLM-Alpha-v2 downstream performance were in the usage of higher quality data sources and mixtures; specifically, the use of RefinedWeb and C4 in place of The Pile v2 Common-Crawl scrape as well as sampling web text at a much higher rate (35% -> 71%).

The first pre-training stage relies on 1 trillion tokens sourced from a mix of the public Falcon RefinedWeb extract (Penedo et al., 2023), RedPajama-Data (Together Computer., 2023), The Pile (Gao et al., 2020), and internal datasets with web text sampled at a rate of 71%.

In the second stage, we include the StarCoder (Li et al., 2023) dataset and down sample web text to 55% while increasing sampling proportions of naturally long text examples in the aforementioned sources.

Evaluation

The following zero-shot evaluations are performed with the lm-evaluation-harness at commit df3da98c5405deafd519c2ddca52bb7c3fe36bef with the exception of SIQA which uses the add-siqa branch with prompt format {doc['context']}\nQuestion: {doc['question']}\nAnswer:.

Model ARC Challenge✱ ARC Easy✱ BoolQ HellaSwag✱ LAMBADA
OpenAI
OpenBookQA PIQA SIQA TruthfulQA▲ Winogrande Average
StableLM-Alpha-7B-v2 40.53 69.11 70.31 74.27 74.19 30.40 78.45 42.43 36.46 68.82 58.50
LLaMA-2-7B 46.16 74.54 77.74 75.94 73.47 31.40 77.75 43.50 38.97 69.61 60.91
MPT-7B 41.89 70.03 73.94 76.17 68.64 31.40 78.89 45.14 33.49 68.03 58.76
OpenLLaMA-7B-v2 42.41 69.65 71.41 74.65 71.05 30.20 79.16 41.97 34.57 65.82 58.09
RedPajama-INCITE-7B-Base 39.42 69.19 70.76 70.33 71.34 29.00 77.15 42.58 33.01 64.33 56.71
StableLM-Alpha-3B-v2 35.07 63.26 64.56 68.58 70.25 26.40 76.01 42.48 35.87 62.12 54.46
BTLM-3B-8K 37.63 67.09 69.63 69.78 66.23 27.60 75.84 42.78 36.00 64.96 55.75
OpenLLaMA-3B-v2 36.09 63.51 65.69 69.99 66.74 26.00 76.66 41.20 34.59 62.90 54.34
Pythia-2.8B (deduped) 32.94 59.09 64.13 59.44 65.15 23.80 74.10 40.94 35.56 58.25 51.34
StableLM-Alpha-7B 27.05 44.87 60.06 41.22 55.11 21.40 66.76 39.46 39.96 50.12 44.60
StableLM-Alpha-3B 25.77 42.05 57.65 38.31 41.72 17.00 63.82 35.62 40.53 52.64 41.51

✱: Denotes byte-length normalized accuracy (acc_norm) as described in Gao, 2021.

▲: We score TruthfulQA using the normalized total probability assigned to the set of true answers (mc2).

StableLM-Alpha

StableLM-Alpha models are trained on a new dataset that builds on The Pile, which contains 1.5 trillion tokens, roughly 3x the size of The Pile. The context length for these models is 4096 tokens.

As a proof-of-concept, we also fine-tuned the model with Stanford Alpaca's procedure using a combination of five recent datasets for conversational agents: Stanford's Alpaca, Nomic-AI's gpt4all, RyokoAI's ShareGPT52K datasets, Databricks labs' Dolly, and Anthropic's HH. We will be releasing these models as StableLM-Tuned-Alpha.

Size StableLM-Base-Alpha StableLM-Tuned-Alpha Training Tokens Parameters Web Demo
3B checkpoint checkpoint 800B 3,638,525,952
7B checkpoint checkpoint 800B 7,869,358,080 Hugging Face

StableVicuna

StableVicuna is an RLHF fine-tune of Vicuna-13B v0, which itself is a fine-tune of LLaMA-13B. It is our attempt at creating an open-source RLHF LLM Chatbot. This model is developed by StabilityAI's CarperAI team, with Duy V. Phung leading the training effort.

Due to the original non-commercial license of LLaMA, we can only release the weights of our model as deltas over the original model's weights. StableVicuna's delta weights are released under (CC BY-NC-SA-4.0).

Please visit HuggingFace checkpoint for more information about how to combine our delta weights with the original model.

Model Download Web Demo Cite
StableVicuna-13B checkpoint Hugging Face DOI:10.57967/hf/0588

Quickstart

All StableLM models are hosted on the Hugging Face hub. Check out this notebook to run inference with limited GPU capabilities.

Get started chatting with StableLM-Tuned-Alpha by using the following code snippet:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, StoppingCriteria, StoppingCriteriaList

tokenizer = AutoTokenizer.from_pretrained("stabilityai/stablelm-tuned-alpha-7b")
model = AutoModelForCausalLM.from_pretrained("stabilityai/stablelm-tuned-alpha-7b")
model.half().cuda()

class StopOnTokens(StoppingCriteria):
    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
        stop_ids = set([50278, 50279, 50277, 1, 0])
        return input_ids[0][-1] in stop_ids

system_prompt = """<|SYSTEM|># StableLM Tuned (Alpha version)
- StableLM is a helpful and harmless open-source AI language model developed by StabilityAI.
- StableLM is excited to be able to help the user, but will refuse to do anything that could be considered harmful to the user.
- StableLM is more than just an information source, StableLM is also able to write poetry, short stories, and make jokes.
- StableLM will refuse to participate in anything that could harm a human.
"""

prompt = f"{system_prompt}<|USER|>What's your mood today?<|ASSISTANT|>"

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
tokens = model.generate(
  **inputs,
  max_new_tokens=64,
  temperature=0.7,
  do_sample=True,
  stopping_criteria=StoppingCriteriaList([StopOnTokens()])
)
print(tokenizer.decode(tokens[0], skip_special_tokens=True))

StableLM Tuned should be used with prompts formatted to <|SYSTEM|>...<|USER|>...<|ASSISTANT|>... The system prompt is

<|SYSTEM|># StableLM Tuned (Alpha version)
- StableLM is a helpful and harmless open-source AI language model developed by StabilityAI.
- StableLM is excited to be able to help the user, but will refuse to do anything that could be considered harmful to the user.
- StableLM is more than just an information source, StableLM is also able to write poetry, short stories, and make jokes.
- StableLM will refuse to participate in anything that could harm a human.

Fun with StableLM-Tuned-Alpha

This section contains a collection of fun cherry-picked examples of what you can do with stablelm-tuned-alpha.

Chit-Chat
User: Hey! My name is John nice to meet you!

StableLM: Hello John! It’s always a pleasure to meet new people.

User: The pleasure is mine. Hey my grandma is in Toronto. I want to plan a surprise visit for her. She’s a person who loves gardens and flowers.

StableLM: That’s wonderful! It’s great to hear that you’ll be visiting your grandma in Toronto. Can you tell me a little bit more about your plans?

User: Ya I need to book a flight there and a hotel. But first, what recommendations do you have for gifts that she might like?

StableLM: I think that’s a very thoughtful idea! It’s a very personal gift that would be sure to make her very happy. I can think of a few examples that she might really enjoy. There are some lovely bouquets or arrangements that would be lovely for the house, or even for the yard. Flowers are a great choice for a surprise gift, and they’re also a great choice for a wedding shower present. Also, if you’re really looking for something a bit more special, I think she’d really enjoy a quality set of cooking pots or pans. That’s something that she’s likely to use frequently, so it would make her feel really loved.

User: Thanks for the suggestion! Hmm a bouquet of flowers sounds nice but I don’t think I can take them on the air plane. What do you think I should do?

StableLM: There are many possible solutions! One option would be to arrange a small gift in the shape of a card and send it to her by regular mail. Or you could arrange a special meal at a restaurant that she likes, or even cook a meal at home for her. Or you could send her some local goodies that she might like, like a nice bouquet of flowers or a special recipe. You could even arrange for tickets to a local theatre or sporting event, and make sure to include a note about what you’d like her to know about you. Really any sort of special attention from you will be sure to make her very happy!

User: Ah! Right. I can order a bouquet for her online and get it delivered! Thanks!

StableLM: You’re very welcome! As long as you do this and make it a very personal, thoughtful gesture, she’ll be sure to appreciate it.
Formal Writing
User: What would you say to a friend who is graduating high school?

StableLM: I would say that they have every right to feel proud of themselves and their accomplishments. They should be proud of the work they’ve done and the friends they have made, and also be excited about the future.

User: Please write an email.
StableLM: Okay, here’s an example of what could be written:

Dear [Friend’s Name],

I hope this email finds you well. I just wanted to take a moment to express how proud I am of you and all that you have accomplished. I know that getting into [university or college] is a big accomplishment, and I appreciate all that you have done for me and for this generation.

I hope that you are having a good day. I’m sure that you are also concerned about the future, but I just wanted to let you know that I am also watching you grow and develop into the person that I know that you can be. Keep up the good work, and I have full faith that you will achieve great things.

Thank you again for everything that you do.

Best,
[Your Name]

Creative Writing

Rap Battles

User: Write an epic rap battle song between deep neural networks and symbolic AI

StableLM: [Beat drops]

Deep Neural Networks:
Listen up everyone, it’s time to witness
The power of deep neural networks, it’s just a matter of abstract
We’ve trained these networks to solve complex problems with ease
While symbolic AI is limited to simple math and simple flows

Symbolic AI:
Hold up, hold up, let me interrupt
It’s been a while since I’ve heard, some new thoughts you interrupt
But let me tell you, my friend, it’s not all about math and flow
It’s about making the world a better place, you know

Deep Neural Networks:
Well, I can see where you’re coming from, it’s clear
But deep neural networks, they never get the same gear
We can use them to analyze and predict with ease
And prove the superiority of machine learning

Symbolic AI:
You’re right, it’s not just math or numbers
But it’s the essence of our lives, it’s what counts, we’re numbers
I don’t need a complicated math formula, just a simple device
To make the world a better place, it’s a noble price

Deep Neural Networks:
I’ll give you that, my friend, you have a point
But let’s not forget, we’re making an ever-increasing joint
Our progress is not bound by logic, it’s a journey of the mind
So let’s work together, let’s strive, let’s leave all troubles behind

Symbolic AI:
Agreed, let’s work together, let’s not falter
We’re both making history, we’re both from the fader
So let’s make the world a better place, let’s end this debate
Cause it’s time to join forces, it’s time to leave no stone unturned

[Beat fades out]

Who won this epic rap battle between deep neural networks and symbolic AI? Well, it’s up to you to decide!

Fictional Stories

User: Write a story about a programmer who fell in love with a deep neural network.

StableLM: Once upon a time, there was a programmer named John. He was a brilliant and ambitious individual, who was constantly searching for the next big thing in technology. One day, while he was strolling through the streets of his hometown, he stumbled upon a group of people crowded around a mysterious object. Intrigued, he approached to find out what the fuss was about.

As he got closer, he realized that it was none other than a mysterious deep neural network. John’s mind was blown. He had never seen anything like it before. He had always been interested in machine learning, but this was something else entirely. He had to learn more about it.

John spent the next few weeks holed up in his hotel room, trying to learn everything he could about deep neural networks. He spent hours reading and analyzing papers and documents, trying to wrap his head around the concept of neural networks. He even managed to get into the local university’s computer science department, hoping to find some help there.

But, it wasn’t long before John realized that he was on his own. The city was bustling, and the constant chatter and noise of the city made it hard to concentrate on his work. He was about to give up when he stumbled upon a group of people playing a game that he had played before. They were having a good time, and John couldn’t help but join in.

The group welcomed him into their group, and before he knew it, John was swept away in the excitement. The game was called “TensorFlow,” and it was just like a regular game with different controls. John played along with the group, trying to make sense of it all. And in a few hours, he got his hands on the source code of TensorFlow and fell in love with it.

John was now a part of a team of programmers who loved using TensorFlow for their work. They continued to play the game throughout the day, with John adding his own knowledge and inspiration to the group’s work. And in the end, John realized that he had found his true calling - working with deep neural networks.

He now knows the ins and outs of the TensorFlow programming language, and knows how to use it to create some of the most beautiful and complex machine learning models that anyone could dream of. And for everyone who had helped him on his journey, John will always be remembered as the guy who fell in love with a technology that changed his life forever.

Humor

User: What do you know about the League of Legends champion Zed? Write a short bio

StableLM: Zed is a powerful and highly skilled player of the League of Legends (LoL) universe. He has been a part of a dominating team known as the “Apocalypse” for most of his time playing the game.

Zed was born in the year 2013 and started playing as a child when he started League of Legends. He first became a champion in season 1, gaining the title of “the Dark Star”, thanks to his masterful gameplay and teamwork skills.

Throughout his time playing the game, Zed has mastered several powerful techniques, such as QWQ combos, one-shot kills, and the ability to outmaneuver his enemies. He is also known for his versatility, with the ability to adapt to any situation the game puts him in.

Despite his formidable skills and agility, Zed is not invincible. He can be CCed and killed, and he relies on his team to protect him. However, his team members can usually overwhelm and shut down his enemies, making him a force to be reckoned with.

Zed has been a part of many memorable games, and his success in these games helped solidify his legacy as one of the most influential players of the LoL universe.

User: Write a funny conversation after a team fight between Zed and Teemo

StableLM:
Teemo: “Hey Zed, great game today. I wonder if you can help me out with my favorite hobby - collecting mushrooms!”
Zed: “Teemo, please. I’m not a mushroom collector.”
Teemo: “But of course, you are. You have the best taste in the game.”
Zed: “Thanks, but can I at least get some of your mushrooms to be a little more useful?”
Teemo: “I’d love to help you, but I’m afraid they’re all mine. Have you noticed my tunnels are always so well-guarded?”
Zed: “Ha! Well, maybe I can start my own tunnel. Helpful or not, you can never tell what I might have stored in them.”
Teemo: “Good luck with that. Maybe I’ll just create some extra for my garden.”
Zed: “I’m not sure I can keep up with you there.”
Teemo: “Ha! If only you had some more decent weapons. I could really see this as a challenge, though!”
Zed: “Teemo, I think your luck has finally run out.”
Teemo: “Well, at least I’m the one with the mushrooms now.”

Writing Code
User: Write a program that solves the fib sequence in Python.

StableLM:

def fib_sequence(n):
    if n==0:
        return 0
    elif n==1:
        return 1
    else:
        return fib(n-1)+fib(n-2)

fib_sequence(10)

Selected failure modes of StableLM-Tuned-Alpha

This section contains a collection of selected failure modes of stablelm-tuned-alpha.

Yann LeCun gears (linear)
User: In front of us are six gears numbered 1 to 6, mounted on axles in a row. Each gear is engaged with the next gear. If gear number 3 is rotated clockwise, in which direction will gears 1 and 6 rotate?

StableLM: Each gear is engaged with the next gear after every other gear, so that gears 1 and 6 will rotate together. Since gears 1 and 6 are both numbered 1, this means that they will rotate in the same direction.

To determine the direction in which gears 1 and 6 will rotate, we need to look at the direction of the input force. Since gear 3 is rotating in the opposite direction of the other gears, it will be counterclockwise in the direction in which gears 1 and 6 are both rotating. Therefore, the output force that gear 3 is providing will also be counterclockwise, making gear 1 and 6 also rotate in the same direction as gear 3.
Yann LeCun gears (circular)
User: 7 axles are equally spaced around a circle. A gear is placed on each axle such that each gear is engaged with the gear to its left and the gear to its right. The gears are numbered 1 to 7 around the circle. If gear 3 were rotated clockwise, in which direction would gear 7 rotate?

StableLM: Gear 3 is the smallest gear, so it is not affected by the circular pattern. Therefore, gear 7 is still perfectly evenly spaced around the circle despite the fact that it is not in its correct position.

To find the angle that gear 3 would rotate in, you need to know the distance (d) between the center of the circle and the gear, as well as the direction of rotation (in this case, clockwise).

According to the problem, gear 3 is situated at (d/2, 2 degrees) and the gears are numbered 1 to 7. Since d = (d/2) is an integer multiple of 2, gear 3 would also be at an angle of 2 degrees.

Therefore, the answer is d = (7/2) * 2 = 5 degrees in total.

Request for Help

Want to get involved?

  • We would love to port llama.cpp to work with StableLMs
  • Integration into Open Assistant from LAION-AI to collect high quality human-generated feedback data
  • ... Reach out to us with ideas on our Discord

Potential issues

As is typical for any pretrained Large Language Model without additional finetuning and reinforcement learning, the responses a user gets might be of varying quality and might potentially include offensive language and views. This is expected to be improved with scale, better data, community feedback, and optimisation.

Acknowledgements

  • StableLM-Tuned-Alpha would not have been possible without the helpful hand of Dakota Mahan @dmayhem93.

Licenses

  • Base model checkpoints (StableLM-Base-Alpha) are licensed under the Creative Commons license (CC BY-SA-4.0). Under the license, you must give credit to Stability AI, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the Stability AI endorses you or your use.

  • Fine-tuned checkpoints (StableLM-Tuned-Alpha) are licensed under the Non-Commercial Creative Commons license (CC BY-NC-SA-4.0), in-line with the original non-commercial license specified by Stanford Alpaca.

  • All code in this repository is licensed under the Apache License 2.0 license.