|
|
--- |
|
|
license: apache-2.0 |
|
|
language: |
|
|
- en |
|
|
tags: |
|
|
- text-to-sql |
|
|
- text to sql |
|
|
pretty_name: Presto/Athena Text to SQL Dataset |
|
|
size_categories: |
|
|
- 1K<n<10K |
|
|
--- |
|
|
|
|
|
I created this dataset using [sqlglot](https://github.com/tobymao/sqlglot) to auto-convert the Spider and Wikisql datasets to Presto syntax, along with running some regex's for additional cleanup. |
|
|
|
|
|
An example use case is fine-tuning an existing model to respond with Presto/Athena text-to-sql, if it performs well at standard SQL syntax used by the major text to sql training datasets. |
|
|
|
|
|
Example of fine-tuning using this dataset (in this case for Mystral 7b Instruct): |
|
|
|
|
|
``` |
|
|
import json |
|
|
import pandas as pd |
|
|
from datasets import Dataset |
|
|
|
|
|
def read_jsonl(file_path): |
|
|
data = [] |
|
|
with open(file_path, 'r', encoding='utf-8') as file: |
|
|
for line in file: |
|
|
json_data = json.loads(line.strip()) |
|
|
data.append(json_data) |
|
|
return data |
|
|
|
|
|
# Read the train and validation files |
|
|
train_data = read_jsonl('training_data/train.jsonl') # use your own path to the training/validation data here |
|
|
valid_data = read_jsonl('training_data/valid.jsonl') |
|
|
|
|
|
# Convert to pandas DataFrame |
|
|
train_df = pd.DataFrame(train_data) |
|
|
valid_df = pd.DataFrame(valid_data) |
|
|
|
|
|
# Convert DataFrame to Huggingface Dataset |
|
|
train_dataset = Dataset.from_pandas(train_df) |
|
|
valid_dataset = Dataset.from_pandas(valid_df) |
|
|
|
|
|
# Example of processing |
|
|
# train_texts = [example['text'] for example in train_dataset] |
|
|
# valid_texts = [example['text'] for example in valid_dataset] |
|
|
|
|
|
instruct_tune_dataset = { |
|
|
"train": train_dataset, |
|
|
"test": valid_dataset |
|
|
} |
|
|
|
|
|
... |
|
|
|
|
|
def create_prompt(sample): |
|
|
""" |
|
|
Update the prompt template: |
|
|
Combine both the prompt and input into a single column. |
|
|
|
|
|
""" |
|
|
bos_token = "<s>" |
|
|
original_system_message = "table:" |
|
|
system_message = "Use the provided input to create an instruction that could have been used to generate the response with an LLM. The query dialect is Athena/Presto. The database table used is: " |
|
|
question_and_response = sample["text"].replace(original_system_message, "").replace("Q:", "\n\n### Input:\n").replace("A:","\n### Response:\n").strip() |
|
|
eos_token = "</s>" |
|
|
|
|
|
full_prompt = "" |
|
|
full_prompt += bos_token |
|
|
full_prompt += "### Instruction:" |
|
|
full_prompt += "\n" + system_message |
|
|
full_prompt += "\n" + question_and_response |
|
|
full_prompt += eos_token |
|
|
|
|
|
return full_prompt |
|
|
|
|
|
... |
|
|
|
|
|
from trl import SFTTrainer |
|
|
|
|
|
trainer = SFTTrainer( |
|
|
model=model, |
|
|
peft_config=peft_config, |
|
|
max_seq_length=max_seq_length, |
|
|
tokenizer=tokenizer, |
|
|
packing=True, |
|
|
formatting_func=create_prompt, # this will apply the create_prompt mapping to all training and test dataset |
|
|
args=args, |
|
|
train_dataset=instruct_tune_dataset["train"], |
|
|
eval_dataset=instruct_tune_dataset["test"] |
|
|
) |
|
|
``` |
