metadata

license: llama2
library_name: peft
tags:
  - llama-2
datasets:
  - ehartford/dolphin
inference: false
pipeline_tag: text-generation
base_model: meta-llama/Llama-2-13b-hf

llama-2-13b-dolphin 🦙🐬

This instruction model was built via parameter-efficient QLoRA finetuning of llama-2-13b on the first 100k rows of ehartford/dolphin (an open-source implementation of Microsoft's Orca). Finetuning was executed on a single A6000 (48 GB) for roughly 18 hours on the Lambda Labs platform.

Benchmark metrics

Metric	Value
MMLU (5-shot)	55.76
ARC (25-shot)	59.98
HellaSwag (10-shot)	82.43
TruthfulQA (0-shot)	42.59
Avg.	60.19

We use state-of-the-art Language Model Evaluation Harness to run the benchmark tests above, using the same version as Hugging Face's Open LLM Leaderboard.

Helpful Links

Model license: Llama 2 Community License Agreement
Basic usage: notebook
Finetuning script: script
Loss curves: plot
Runtime stats: table

Example prompts and responses

Example 1:

User:

You are a helpful assistant. Write me a numbered list of things to do in New York City.\n

llama-2-13b-dolphin-peft:

Visit the iconic Statue of Liberty and Ellis Island.

Take a stroll through Central Park and enjoy its many attractions.

Explore the world-renowned museums, such as the Metropolitan Museum of Art and the Museum of Modern Art.

Experience the vibrant energy of Times Square and take in the bright lights and billboards.

Visit the 9/11 Memorial and Museum to pay tribute to those who lost their lives in the attacks.

Enjoy a Broadway show or a concert at one of the many theaters and venues in the city.

Take a ride on the Staten Island Ferry for a free view of the Statue of Liberty and the New York City skyline.

Shop at the famous Fifth Avenue stores and explore the high-end boutiques.

Indulge in a variety of cuisines at one of the many restaurants in the city.

Visit the Empire State Building and enjoy the panoramic views of the city from the observation deck.

Example 2:

User:

You are a helpful assistant. Write a short email inviting my friends to a dinner party on Friday. Respond succinctly.\n"

llama-2-13b-dolphin-peft:

Subject: Friday Dinner Party Invitation

Dear Friends,

I hope this email finds you well. I'm excited to invite you all to a dinner party on Friday, March 10th, at 7:00 PM. The address is 123 Main Street, Anytown, USA.

Please RSVP by Wednesday, March 8th, so I can plan accordingly. I look forward to seeing you all and sharing a delicious meal together!

Best, Your Friendly Assistant

P.S. If you have any dietary restrictions or allergies, please let me know in your RSVP. Thank you!

Model Description

The architecture is a modification of a standard decoder-only transformer.

The llama-2-13b models have been modified from a standard transformer in the following ways:

It uses the SwiGLU activation function
It uses rotary positional embeddings (RoPE)

Hyperparameter	Value
n_parameters	13B
tokens	2.0T
vocab size	32000
sequence length	4096

Finetuning Description

This model was trained on a single A6000 (48 GB) for about 18 hours using the Lambda Labs platform.

The above loss curve was generated from the run's private wandb.ai log.

PreTraining Data

For more details on the pretraining process, see Llama-2-13b-hf.

The data was tokenized using the Llama-2-13b-hf tokenizer.

Limitations and Biases

The following language is modified from EleutherAI's GPT-NeoX-20B

This model can produce factually incorrect output, and should not be relied on to produce factually accurate information. This model was trained on various public datasets. While great efforts have been taken to clean the pretraining data, it is possible that this model could generate lewd, biased or otherwise offensive outputs.

How to Use

Basic usage: notebook

Install and import the package dependencies:

!pip install -q -U huggingface_hub peft transformers torch accelerate

import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

Sign into a HF account with access to Llama-2:

from huggingface_hub import notebook_login
notebook_login()

Basic model loading:

peft_model_id = "dfurman/llama-2-13b-dolphin-peft"
config = PeftConfig.from_pretrained(peft_model_id)

tokenizer = AutoTokenizer.from_pretrained(
    config.base_model_name_or_path,
    use_auth_token=True
)
tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForCausalLM.from_pretrained(
    config.base_model_name_or_path,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    use_auth_token=True,
)

# Load the Lora model
model = PeftModel.from_pretrained(model, peft_model_id)

Once loaded, the model and tokenizer can be used with the following code:

def llama_generate(
    model: AutoModelForCausalLM,
    tokenizer: AutoTokenizer,
    prompt: str,
    max_new_tokens: int = 128,
    temperature: float = 0.92,
) -> str:
    """
    Initialize the pipeline
    Uses Hugging Face GenerationConfig defaults
        https://huggingface.co/docs/transformers/v4.29.1/en/main_classes/text_generation#transformers.GenerationConfig
    Args:
        model (transformers.AutoModelForCausalLM): Falcon model for text generation
        tokenizer (transformers.AutoTokenizer): Tokenizer for model
        prompt (str): Prompt for text generation
        max_new_tokens (int, optional): Max new tokens after the prompt to generate. Defaults to 128.
        temperature (float, optional): The value used to modulate the next token probabilities.
            Defaults to 1.0
    """
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    inputs = tokenizer(
        [prompt],
        return_tensors="pt",
        return_token_type_ids=False,
    ).to(
        device
    )  # tokenize inputs, load on device

    # when running Torch modules in lower precision, it is best practice to use the torch.autocast context manager.
    with torch.autocast("cuda", dtype=torch.bfloat16):
        response = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            temperature=temperature,
            return_dict_in_generate=True,
            eos_token_id=tokenizer.eos_token_id,
            pad_token_id=tokenizer.pad_token_id,
        )

    decoded_output = tokenizer.decode(
        response["sequences"][0],
        skip_special_tokens=True,
    )  # grab output in natural language

    return decoded_output[len(prompt) :]  # remove prompt from output

We can now generate text! For example:

prompt = "### Human: Write me a numbered list of things to do in New York City.### Assistant: "

response = llama_generate(
    model,
    tokenizer,
    prompt,
    max_new_tokens=250,
    temperature=0.92,
)

print(response)

Runtime tests

runtime / 50 tokens (sec)	GPU	attn	torch dtype	VRAM (GB)
2.93	1x A100 (40 GB SXM)	torch	bfloat16	25
3.24	1x A6000 (48 GB)	torch	bfloat16	25

The above runtime stats were generated from this notebook.

Acknowledgements

This model was finetuned by Daniel Furman on July 22, 2023 and is intended primarily for research purposes.

Disclaimer

The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please cosult an attorney before using this model for commercial purposes.

Meta citation for llama-2 blog

@online{Meta2023Introducing,
    author    = {Meta AI},
    title     = {Meta and Microsoft Introduce the Next Generation of Llama},
    year      = {2023},
    url       = {https://about.fb.com/news/2023/07/llama-2/},
    note      = {Accessed: 2023-07-24},
    urldate   = {2023-07-24}
}

Framework versions

PEFT 0.5.0.dev0