src/data_utils.py

from torch.nn.functional import pad
from torch.utils.data import Dataset
import torch
import json
from transformers import PreTrainedTokenizer

from dataclasses import dataclass


def longest_common_subsequence(a, b, s_i=0, s_j=0) -> list:
    a = a.numpy()
    b = b.numpy()
    m, n = len(a), len(b)
    
    i = s_i
    j = s_j
    result = []

    while i < m and j < n:
        if a[i][1] == 0:
            i += 1
            continue
        if b[j] == 0:
            j += 1
            continue
            
        if a[i][1] == b[j]:
            result.append(i+1)
            i += 1
            j += 1
        elif a[i][1] < b[j]:
            i += 1
        else:
            j += 1
            
    return result

def get_pooler_tensor(segments_idxs):
    # Tạo chỉ số segment đã pad cho toàn bộ batch
    padded_idx_batch = []
    max_seg, max_len_all = 0, 0
    pad_multiple = 4

    for seg_idx, max_len in segments_idxs:
        max_len_all = max(max_len_all, max_len)
        max_seg = max(max_seg, len(seg_idx))

        padded = torch.stack([
            pad(x, (0, max_len - len(x)), value=-1)
            for x in seg_idx
        ])  # (num_segments, max_len)

        padded_idx_batch.append(padded)

    # Pad toàn bộ batch về cùng shape (B, max_seg, max_len_all)
    def pad2d(t, h, w):
        return pad(t, (0, w - t.size(1), 0, h - t.size(0)), value=-1)

    # max_seg = int(math.ceil(max_seg / pad_multiple) * pad_multiple)
    padded_idx_batch = torch.stack([
        pad2d(p, max_seg, max_len_all) for p in padded_idx_batch
    ])  # (B, max_seg, max_len_all)

    # Tạo mask và gather từ X
    mask = padded_idx_batch != -1
    safe_idx = padded_idx_batch.masked_fill(~mask, 0)

    return {'safe_idx': safe_idx, 'mask': mask}

def prepare_pooler(offset_mapping, starts, phrases_offsets):
    seg_idxs = []
    for offset, start, phrases_offset in zip(offset_mapping, starts, phrases_offsets):

        seg_idx = []
       
        token_offset_start = [start.item()]

        longest_common_offset = token_offset_start + longest_common_subsequence(offset, phrases_offset, start) 
        student_max_len = 1

        for i in range(1, len(longest_common_offset)):
            seg_idx.append(torch.arange(longest_common_offset[i - 1], longest_common_offset[i]))
            student_max_len = max(student_max_len, seg_idx[-1].size(0))

        seg_idxs.append((seg_idx, student_max_len))

    return get_pooler_tensor(seg_idxs)


class LLMDataset(Dataset):
    def __init__(self, file_path, tokenizer, syntactic_file, prompt_max_len=512):

        self.dataset = []

        with open(file_path, "r", encoding="utf-8") as f:
            for line in f:
                data = json.loads(line)
                self.dataset.append(data)

                s_prompt = tokenizer(
                    data['prompt'], 
                    max_length=prompt_max_len,
                    truncation=True, 
                    add_special_tokens=False
                )
                data['prompt'] = tokenizer.decode(s_prompt['input_ids'])
                data['prompt_len'] = len(s_prompt['input_ids'])
        
        with open(syntactic_file, "r", encoding="utf-8") as f:
            idx = 0
            for line in f:
                prompt_end = len(self.dataset[idx]['prompt'])
                data = json.loads(line)
                phrases_lvl1 = [prompt_end] + [prompt_end + item['end_char'] for item in data['phrases_lvl1']]
                phrases_lvl2 = [prompt_end] + [prompt_end + item['end_char'] for item in data['phrases_lvl2']]

                self.dataset[idx]['phrases_lvl1_offset'] = torch.tensor(phrases_lvl1)
                self.dataset[idx]['phrases_lvl2_offset'] = torch.tensor(phrases_lvl2)

                idx += 1


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

    def __getitem__(self, index):
        return (self.dataset[index]['prompt'], 
                self.dataset[index]['output'], 
                self.dataset[index]['prompt_len'],
                self.dataset[index]['phrases_lvl1_offset'], 
                self.dataset[index]['phrases_lvl2_offset'])
    

@dataclass
class LLMDataCollator:
    tokenizer: PreTrainedTokenizer = None
    model_type: str = ''
    do_train: bool = True
    max_len: int = 512
    pad_to_multiple_of: int = 4
    return_tensors: str = 'pt'
    padding: bool = True
    return_offsets_mapping: bool = True
    n_span: int = 4


    def __call__(self, batch):
        prompts, fulls, prompt_lengths, phrases_lvl1_offsets, phrases_lvl2_offsets = [], [], [], [], []
        for prompt, output, prompt_length, phrases_lvl1_offset, phrases_lvl2_offset in batch:
            prompts.append(prompt)
            fulls.append(prompt + output)
            prompt_lengths.append(prompt_length)
            phrases_lvl1_offsets.append(phrases_lvl1_offset)
            phrases_lvl2_offsets.append(phrases_lvl2_offset)

        
        inputs = self.tokenizer(
            fulls,
            truncation=True,
            padding=self.padding,
            max_length=self.max_len - 1,
            return_tensors=self.return_tensors,
            pad_to_multiple_of=self.pad_to_multiple_of,
            return_offsets_mapping=self.return_offsets_mapping and self.do_train,
            add_special_tokens=False
        )
        

        eos_tokens = torch.full((inputs["input_ids"].size(0), 1), self.tokenizer.eos_token_id, dtype=torch.long)
        inputs["input_ids"] = torch.cat([inputs["input_ids"], eos_tokens], dim=1)
        inputs["attention_mask"] = torch.cat([inputs["attention_mask"], 
                                              torch.zeros((inputs["attention_mask"].size(0), 1), dtype=torch.long)], dim=1)
        
        labels = inputs["input_ids"][:, 1:].clone().detach()
        labels = torch.cat([labels, torch.full((labels.size(0), 1), -100, dtype=torch.long)], dim=1)

        input_lengths = inputs["attention_mask"].sum(dim=1)
        prompt_lengths = torch.tensor(prompt_lengths)

        if self.model_type in ["gpt2"]:
            position_ids = torch.zeros(inputs['input_ids'].size(), dtype=torch.long)
            for i in range(input_lengths.size(0)):
                position_ids[i, :input_lengths[i]] = torch.arange(0, input_lengths[i], dtype=torch.long)
            inputs["position_ids"] = position_ids

        for i in range(len(labels)):
            labels[i, :(prompt_lengths[i] -1)] = -100
            labels[i, input_lengths[i]:] = -100

        if not self.do_train:
            return inputs, None, labels

        token_offset_mapping = inputs.pop('offset_mapping', None)
        if token_offset_mapping is not None :
            starts = torch.zeros_like(prompt_lengths)
            pooler_tensor = prepare_pooler(token_offset_mapping, starts, phrases_lvl2_offsets)

            inputs['pooler_safe_idx'] = pooler_tensor['safe_idx']
            inputs['pooler_mask'] = pooler_tensor['mask']

        return inputs, labels