code/utils/load_sft_dataset.py

import json
import torch
import gc

from datasets import Dataset
from torch.utils.data import Dataset
from schema_item_filter import SchemaItemClassifierInference, filter_schema, filter_schema_purple
from utils.db_utils import get_db_schema_sequence, get_matched_content_sequence

def prepare_text2sql_prefix_sequence(data):
    prompt = f"""Convert the question to SQL query.
{data["schema_sequence"]}
{data["content_sequence"]}
question: {data["text"]}"""
    return prompt

def prepare_inputs_and_labels(prefix_seq, target_seq, tokenizer, max_tokens):
    prefix_ids = [tokenizer.bos_token_id] + tokenizer(prefix_seq , truncation = False)["input_ids"]
    target_ids = tokenizer(target_seq, truncation = False)["input_ids"] + [tokenizer.eos_token_id]

    seq_length = len(prefix_ids) + len(target_ids)
    if seq_length <= max_tokens: # pad inputs with pad_token_id
        pad_length = max_tokens - seq_length
        input_ids = prefix_ids + target_ids + [tokenizer.pad_token_id] * pad_length
        # tell the model to ignore the padding tokens when performing (masked) self-attention 
        attention_mask = [1] * seq_length + [0] * pad_length
        # only target_ids produces gradients
        labels = [-100] * len(prefix_ids) + target_ids + [-100] * pad_length
    else: # no padding
        print("the current input sequence exceeds the max_tokens, we will truncate it.")
        input_ids = prefix_ids + target_ids
        # pre-truncate input ids
        input_ids = [tokenizer.bos_token_id] + input_ids[-(max_tokens-1):]
        attention_mask = [1] * max_tokens
        # only target_ids produces gradients
        labels = [-100] * len(prefix_ids) + target_ids
        # pre-truncate labels
        labels = labels[-max_tokens:]
    
    return {
        "input_ids": torch.tensor(input_ids, dtype = torch.int64), 
        "attention_mask": torch.tensor(attention_mask, dtype = torch.int64), 
        "labels": torch.tensor(labels, dtype = torch.int64)
    }

# def prepare_inputs(prefix_seq, tokenizer, max_prefix_length):
#     input_ids = [tokenizer.bos_token_id] + tokenizer(prefix_seq , truncation = False)["input_ids"]

#     if len(input_ids) > max_prefix_length:
#         print("the current input sequence exceeds the max_tokens, we will truncate it.")
#         input_ids = [tokenizer.bos_token_id] + input_ids[-(max_prefix_length-1):]
    
#     attention_mask = [1] * len(input_ids)
    
#     return {
#         "input_ids": torch.tensor(input_ids, dtype = torch.int64),
#         "attention_mask": torch.tensor(attention_mask, dtype = torch.int64)
#     }

def prepare_inputs(prefix_seq, tokenizer, max_prefix_length):
    messages =  [{
        'role': 'user',
        'content': prefix_seq
    }]
    prompt = tokenizer.apply_chat_template(messages,  tokenize=False, add_generation_prompt=True)

    input_ids = tokenizer(prompt , truncation = False)["input_ids"]

    if len(input_ids) > max_prefix_length:
        print("the current input sequence exceeds the max_tokens, we will truncate it.")
        input_ids = input_ids[-(max_prefix_length-1):]
    
    attention_mask = [1] * len(input_ids)
    
    return {
        "input_ids": torch.tensor(input_ids, dtype = torch.int64),
        "attention_mask": torch.tensor(attention_mask, dtype = torch.int64)
    }

class SFTSQLGenerationDataset(Dataset):
    def __init__(self, text2sql_data_dir, tokenizer, max_tokens, mode, table_num, column_num, threshold, sic_path, do_filter_schema=True):
        super().__init__()
        dataset = json.load(open(text2sql_data_dir))

        print("apply filtering strategies...")
        if do_filter_schema:
            if mode == "train":
                dataset = filter_schema(dataset, "train", None, table_num, column_num, threshold=threshold)
            elif mode == "eval":
                sic = SchemaItemClassifierInference(sic_path)
                dataset = filter_schema(dataset, "eval", sic, table_num, column_num, threshold=threshold)
                # dataset = filter_schema_purple(dataset, "eval", "/home/datht/llmsql/selector/spider-dev/spider-dev-selector-t0.02-value-samples.json")
                del sic
                torch.cuda.empty_cache()

            # prepare schema sequence and content sequence
            for data in dataset:
                data["schema_sequence"] = get_db_schema_sequence(data["schema"])
                # data["content_sequence"] = get_matched_content_sequence(data["matched_contents"])

        self.mode = mode
        self.dataset = dataset
        self.tokenizer = tokenizer
        self.max_tokens = max_tokens

    def __getitem__(self, index):
        data = self.dataset[index]
        prefix_seq = prepare_text2sql_prefix_sequence(data)
        if index < 2:
            print(prefix_seq)

        if self.mode == "train":
            target_seq = data["sql"]
            return prepare_inputs_and_labels(prefix_seq, target_seq, self.tokenizer, self.max_tokens)
        elif self.mode == "eval":
            return prepare_inputs(prefix_seq, self.tokenizer, self.max_tokens)

    def __len__(self):
        return len(self.dataset)