linhduongtuan/BLOOM-LORA

Why do we try to finetune these BLOOM models? The major reason is the licence of LLaMA, while the BLOOM licence seems to be more relax! Moreover, BLOOM models were trained on the dataset having 59 Languages (46 natural and 13 programing languages instead of around 20 languages in LLaMA. Especially, BLOOM models were trained ond the dataset including 2.7% Vietnamese (~8^th)

We try to reimplement BLOOM-LoRA using a variety of sources such as the original LLaMA, Stanford-Alpaca, Alpaca-LoRA, BLOOMZ, and a name to few. These datasets for finetuning tasks can be found this data folder or my HuggingFace Hub for large ones.

For example, you can try our finetuned BLOOM-b71-alpaca model out on Colab here!!!!!

This repository contains code for reproducing the Stanford Alpaca results using low-rank adaptation (LoRA). We provide an Instruct model of similar quality to text-davinci-003 that can run on a Raspberry Pi (for research), and the code can be easily extended to the 1b1, 3b, 7b1, and 175b models.

In addition to the training code, which runs within five hours on a single RTX 4090, or a bit longer with 3090. we publish a script for downloading and inference on the foundation model and LoRA, as well as the resulting LoRA weights themselves. To fine-tune cheaply and efficiently, we use Hugging Face's PEFT as well as Tim Dettmers' bitsandbytes.

Without hyperparameter tuning or validation-based checkpointing, the LoRA model produces outputs comparable to the Stanford Alpaca model. (Please see the outputs included below.) Further tuning might be able to achieve better performance; I invite interested users to give it a try and report their results.

For discussion and support, users have created a dedicated Twitter server here.

Update 2023-03-27: weights have been updated with the CodeAlpaca-20k dataset from sahil2801/CodeAlpaca-20k. This should intensify this ability of BLOOM-7b1 for generating programming languages, even BLOOM-7b1 has alread trained on datasets containing 13 programming languages. Additionally, we will publicly release our trained weights for these finetuned models at HuggingFace Hub with the addresse 'LinhDuong/bloom-7b1-lora-codealpaca20k'

Update 2023-03-25: weights have been updated with cleaned data and prompts masked out in the loss. This should reduce the number of template artifacts in outputs. Additionally, we will publicly release our trained weights for these finetuned models at HuggingFace Hub with the addresse 'LinhDuong/bloom-7b1-alpaca'

Update 2023-03-21: weights have been updated with cleaned data and prompts masked out in the loss. This should reduce the number of template artifacts in outputs.

1. Install dependencies

pip install -r requirements.txt

2. If bitsandbytes doesn't work, install it from source. Windows users can follow these instructions.

• Please use (train_alpaca.py), if you want to finetune for Bloom-7b1-lora using [alpaca_data_cleaned.json dataset)[https://github.com/gururise/AlpacaDataCleaned/blob/main/alpaca_data_cleaned.json].

• Please use (train_CodeAlpaca20K.py), if you want to finetune Bloom-7b1-lora using (CodeAlpaca-20k dataset)[https://huggingface.co/datasets/sahil2801/CodeAlpaca-20k] for intesifying generative programming languages.

• Otherwise, please use (finetune.py) for the original Alpaca-LoRA, if you want to finetune LLaMA models.

This file contains a straightforward application of PEFT to the BLOOM-7b1 model, as well as some code related to prompt construction and tokenization. Near the top of this file is a set of hardcoded hyperparameters that you should feel free to modify. PRs adapting this code to support larger models are always welcome.

Inference (generate_alpaca.py) for LLaMA ~ Alpaca models

This file reads the foundation model from the Hugging Face model hub and the LoRA weights from tloen/alpaca-lora-7b, and runs a Gradio interface for inference on a specified input. Users should treat this as example code for the use of the model, and modify it as needed.

Otherwise, Inference (generate_bloom.py) for BLOOM-7b1 model

This file reads the foundation model from the Hugging Face model hub and the LoRA weights from LinhDuong/bloom-7b1-alpaca, and runs a Gradio interface for inference on a specified input. Users should treat this as example code for the use of the model, and modify it as needed.

Checkpoint export (export_*_checkpoint.py)

These files contain scripts that merge the LoRA weights back into the base model for export to Hugging Face format and to PyTorch state_dicts. They should help users who want to run inference in projects like llama.cpp or alpaca.cpp.

Dataset

In addition to alpaca_data.json, which contains the original Stanford Alpaca dataset, we also include alpaca_data_cleaned.json, which has been stripped of various tokenization artifacts with the help of @gururise and refer his repository at here. This file is now used by default in the training script.

@AndriyMulyar has also provided interactive, embedding-based visualizations of the original dataset's instructions and outputs, as well as clusters of bad examples.

• We can likely improve our model performance significantly if we combed through the data and fixed bad examples; in fact, dataset quality might be our bottleneck.
• We're continually fixing bugs and conducting training runs, and the weights on the Hugging Face Hub are being updated accordingly. In particular, those facing issues with response lengths should make sure that they have the latest version of the weights and code.

1. I would like to express my gratitude to Phenikaa University and Phenikaa-X company, without whose support, this fine-tuned model would not have been possible. Their generosity in providing me with access to their computation resources, namely A100 and RTX 4090, allowed me to train and test my work successfully.
2. I would also like to extend my thanks to Pham Ngoc Ninh, who worked diligently to address any bugs and refine this model. Ninh's assistance was greatly appreciated.

Instruction: Tell me about alpacas.

BLOOM-7b1-LoRA: Alpacas are native to the Andes Mountains of South America, where they have been domesticated for thousands of years. They are known for their long, fluffy coats, which are used to make clothing, rugs, and other accessories. Alpacas are also known for their gentle temperament, which makes them a popular pet and companion animal.

Instruction: Tell me about the president of Mexico in 2019.

BLOOM-7b1-LoRA: The president of Mexico in 2019 is Andrés Manuel López Obrador. He was elected in 2018 and took office in December of that year. López Obrador is known for his populist style of government and his support for social programs such as free health care and free education. He is also known for his opposition to the United States and its policies in Latin America. López Obrador has also been criticized for his lack of experience in government and his inability to govern effectively. Despite these criticisms, López Obrador has been able to gain the support of the majority of Mexicans and has been able to implement many of his policies.

Instruction: Tell me about the king of France in 2019.

BLOOM-7b1-LoRA: The king of France in 2019 is Emmanuel Macron. Macron was elected as the president of France in May 2017. He is the first president of France to come from outside the traditional political elite. Macron is a former banker and a member of the Socialist Party. He has been a member of the European Parliament since 2014 and has been a member of the Socialist Group in the European Parliament since 2015.

Instruction: List all Canadian provinces in alphabetical order.

BLOOM-7b1-LoRA: Alberta, British Columbia, Manitoba, New Brunswick, Newfoundland and Labrador, Nova Scotia, Northwest Territories, Nunavut, Ontario, Prince Edward Island, Quebec, Saskatchewan, Southwest Territories, Yukon, British Columbia, Alberta, Manitoba, New Brunswick, Newfoundland and Labrador, Nova Scotia, Northwest Territories, Nunavut, Ontario, Prince Edward Island, Quebec, Saskatchewan, Southwest Territories, Yukon, British Columbia, Alberta, Manitoba, New Brunswick, Newfoundland and Labrador, Nova Scotia, Northwest Territories, Nunavut, Ontario, Prince Edward Island, Quebec, Saskatchewan

Instruction: Write a Python program that prints the first 10 Fibonacci numbers.

BLOOM-7b1-LoRA:

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

for i in range(10):
    print(fibonacci(i))

# Output:
# 0
# 1
# 1
# 2
# 3
# 5
# 8
# 13
# 21
# 34

BLOOM-7b1-LoRA finetuned with CodeAlpaca dataset:

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

if __name__ == '__main__':
    for i in range(10):
        print(fibonacci(i))

# Output
# 0
# 1
# 2
# 3
# 5
# 8
# 13
# 21
# 34
# 55
# 89
# 144
# 233
# 377
# 610

a, b = 0, 1
count = 0

while count < 10:
    print(a)
    a, b = b, a + b
    count += 1
0
1
1
2
3
5
8
13
21
34

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

print(fibonacci(10))

numbers = [0, 1]

for i in range(2, 11):
    numbers.append(numbers[i-2] + numbers[i-1])

print(numbers)
# Output: [0, 1, 1, 2, 3, 5, 8, 13, 21, 34]

# Fibonacci numbers are the numbers in the following integer sequence, called the Fibonacci sequence, and characterized by the fact that every number after the first two is the sum of the two preceding ones.

# Fibonacci sequence: 0, 1, 1, 2, 3, 5, 8, 13, 21, 34, ...

# Program:

# Initialize the first two numbers of the Fibonacci sequence
a = 0
b = 1

# Print the first 10 numbers of the Fibonacci sequence
print(a)
print(b)

for i in range(2,10):
    c = a + b
    a = b
    b = c
    print(c)

Instruction: Write a program that prints the numbers from 1 to 100. But for multiples of three print 'Fizz' instead of the number and for the multiples of five print 'Buzz'. For numbers which are multiples of both three and five print 'FizzBuzz'.

BLOOM-7b1-LoRA: This program uses a for loop to iterate over the range of numbers from 1 to 100. For each number, it uses conditional statements (if/elif/else) to determine whether it is a multiple of 3, 5, both, or neither. If it is a multiple of 3, it prints "Fizz". If it is a multiple of 5, it prints "Buzz". If it is a multiple of both 3 and 5, it prints "FizzBuzz". If it is not a multiple of either 3 or 5, it prints the number itself.

def print_numbers(num):
    if num % 3 == 0:
        print('Fizz')
    elif num % 5 == 0:
        print('Buzz')
    else:
        print(num)

print_numbers(1)
print_numbers(2)
print_numbers(3)
print_numbers(4)
print_numbers(5)
print_numbers(6)
print_numbers(7)
print_numbers(8)
print_numbers(9)
print_numbers(10)
print_numbers(11)
print_numbers(12)
print_numbers(13)
print_numbers(14)

BLOOM-7b1-LoRA finetuned with CodeAlpaca dataset: Write a program that prints the numbers from 1 to 100. But for multiples of three print 'Fizz' instead of the number and for the multiples of five print 'Buzz'. For numbers which are multiples of both three and five print 'FizzBuzz'

def print_numbers():
    for i in range(1, 101):
        if i % 3 == 0:
            print('Fizz')
        elif i % 5 == 0:
            print('Buzz')
        else:
            print(str(i))

print_numbers()

# Output:
# Fizz
# Buzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz
# FizzBuzz

for num in range(1, 101):
    if num % 3 == 0 and num % 5 == 0:
        print("FizzBuzz")
    elif num % 3 == 0:
        print("Fizz")
    elif num % 5 == 0:
        print("Buzz")
    else:
        print(num)

for i in range(1, 101):
    if i % 3 == 0 and i % 5 == 0:
        print("FizzBuzz")
    elif i % 3 == 0:
        print("Fizz")
    elif i % 5 == 0:
        print("Buzz")
    else:
        print(i)

for i in range(1, 101):
    if i % 3 == 0 and i % 5 == 0:
        print('FizzBuzz')
    elif i % 3 == 0:
        print('Fizz')
    elif i % 5 == 0:
        print('Buzz')
    else:
        print(i)

for (let i = 1; i <= 100; i++) {
  if (i % 3 === 0 && i % 5 === 0) {
    console.log("FizzBuzz");
  } else if (i % 3 === 0) {
    console.log("Fizz");
  } else if (i % 5 === 0) {
    console.log("Buzz");
  } else {
    console.log(i);
  }
}

Instruction: Tell me five words that rhyme with 'shock'.

BLOOM-7b1-LoRA: The five words that rhyme with "shock" are: Shock, shock, shock, shock, shock

Instruction: Translate the sentence 'I have no mouth but I must scream' into Spanish.

BLOOM-7b1-LoRA: Yo no tengo boca pero debo gritar