Get LLM Finetuning Data

The following are examples of training data in different formats.

Text

from datasets import load_dataset
from superduper.base.document import Document
dataset_name = "timdettmers/openassistant-guanaco"
dataset = load_dataset(dataset_name)

train_dataset = dataset["train"]
eval_dataset = dataset["test"]

train_documents = [
    Document({**example, "_fold": "train"})
    for example in train_dataset
]
eval_documents = [
    Document({**example, "_fold": "valid"})
    for example in eval_dataset
]

datas = train_documents + eval_documents        

Prompt-Response

from datasets import load_dataset
from superduper.base.document import Document
dataset_name = "mosaicml/instruct-v3"
dataset = load_dataset(dataset_name)

train_dataset = dataset["train"]
eval_dataset = dataset["test"]

train_documents = [
    Document({**example, "_fold": "train"})
    for example in train_dataset
]
eval_documents = [
    Document({**example, "_fold": "valid"})
    for example in eval_dataset
]

datas = train_documents + eval_documents        

Chat

from datasets import load_dataset
from superduper.base.document import Document
dataset_name = "philschmid/dolly-15k-oai-style"
dataset = load_dataset(dataset_name)['train'].train_test_split(0.9)

train_dataset = dataset["train"]
eval_dataset = dataset["test"]

train_documents = [
    Document({**example, "_fold": "train"})
    for example in train_dataset
]
eval_documents = [
    Document({**example, "_fold": "valid"})
    for example in eval_dataset
]

datas = train_documents + eval_documents        

We can define different training parameters to handle this type of data.

Text

# Function for transformation after extracting data from the database
transform = None
key = ('text')
training_kwargs=dict(dataset_text_field="text")        

Prompt-Response

# Function for transformation after extracting data from the database
def transform(prompt, response):
    return {'text': prompt + response + "</s>"}

key = ('prompt', 'response')
training_kwargs=dict(dataset_text_field="text")        

Chat

# Function for transformation after extracting data from the database
transform = None

key = ('messages')
training_kwargs=None        

Example input_text and output_text

Text

data = datas[0]
input_text, output_text = data["text"].rsplit("### Assistant: ", maxsplit=1)
input_text += "### Assistant: "
output_text = output_text.rsplit("### Human:")[0]
print("Input: --------------")
print(input_text)
print("Response: --------------")
print(output_text)        

Prompt-Response

data = datas[0]
input_text = data["prompt"]
output_text = data["response"]
print("Input: --------------")
print(input_text)
print("Response: --------------")
print(output_text)        

Chat

data = datas[0]
messages = data["messages"]
input_text = messages[:-1]
output_text = messages[-1]["content"]
print("Input: --------------")
print(input_text)
print("Response: --------------")
print(output_text)