|
|
from typing import TYPE_CHECKING, Any, Dict, List, Optional |
|
|
|
|
|
from ...extras.constants import IGNORE_INDEX |
|
|
from ...extras.logging import get_logger |
|
|
from .mm_utils import get_paligemma_token_type_ids, get_pixel_values |
|
|
|
|
|
|
|
|
if TYPE_CHECKING: |
|
|
from transformers import ProcessorMixin |
|
|
from transformers.tokenization_utils import PreTrainedTokenizer |
|
|
|
|
|
from ...hparams import DataArguments |
|
|
from ..template import Template |
|
|
|
|
|
|
|
|
logger = get_logger(__name__) |
|
|
|
|
|
|
|
|
def preprocess_pairwise_dataset( |
|
|
examples: Dict[str, List[Any]], |
|
|
template: "Template", |
|
|
tokenizer: "PreTrainedTokenizer", |
|
|
processor: Optional["ProcessorMixin"], |
|
|
data_args: "DataArguments", |
|
|
) -> Dict[str, List[List[int]]]: |
|
|
|
|
|
model_inputs = { |
|
|
"chosen_input_ids": [], |
|
|
"chosen_attention_mask": [], |
|
|
"chosen_labels": [], |
|
|
"rejected_input_ids": [], |
|
|
"rejected_attention_mask": [], |
|
|
"rejected_labels": [], |
|
|
} |
|
|
if processor is not None: |
|
|
model_inputs["pixel_values"] = [] |
|
|
if hasattr(processor, "image_seq_length"): |
|
|
model_inputs["chosen_token_type_ids"] = [] |
|
|
model_inputs["rejected_token_type_ids"] = [] |
|
|
|
|
|
for i in range(len(examples["prompt"])): |
|
|
if len(examples["prompt"][i]) % 2 != 1 or len(examples["response"][i]) < 2: |
|
|
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i])) |
|
|
continue |
|
|
|
|
|
if processor is not None and not hasattr(processor, "image_seq_length"): |
|
|
examples["prompt"][i][0]["content"] = template.image_token + examples["prompt"][i][0]["content"] |
|
|
|
|
|
chosen_messages = examples["prompt"][i] + [examples["response"][i][0]] |
|
|
rejected_messages = examples["prompt"][i] + [examples["response"][i][1]] |
|
|
prompt_ids, chosen_ids = template.encode_oneturn( |
|
|
tokenizer, |
|
|
chosen_messages, |
|
|
examples["system"][i], |
|
|
examples["tools"][i], |
|
|
data_args.cutoff_len, |
|
|
data_args.reserved_label_len, |
|
|
) |
|
|
_, rejected_ids = template.encode_oneturn( |
|
|
tokenizer, |
|
|
rejected_messages, |
|
|
examples["system"][i], |
|
|
examples["tools"][i], |
|
|
data_args.cutoff_len, |
|
|
data_args.reserved_label_len, |
|
|
) |
|
|
|
|
|
if template.efficient_eos: |
|
|
chosen_ids += [tokenizer.eos_token_id] |
|
|
rejected_ids += [tokenizer.eos_token_id] |
|
|
|
|
|
if processor is not None and hasattr(processor, "image_seq_length"): |
|
|
image_token_id = tokenizer.convert_tokens_to_ids(template.image_token) |
|
|
prompt_ids = [image_token_id] * getattr(processor, "image_seq_length") + prompt_ids |
|
|
|
|
|
chosen_input_ids = prompt_ids + chosen_ids |
|
|
chosen_labels = [IGNORE_INDEX] * len(prompt_ids) + chosen_ids |
|
|
rejected_input_ids = prompt_ids + rejected_ids |
|
|
rejected_labels = [IGNORE_INDEX] * len(prompt_ids) + rejected_ids |
|
|
model_inputs["chosen_input_ids"].append(chosen_input_ids) |
|
|
model_inputs["chosen_attention_mask"].append([1] * len(chosen_input_ids)) |
|
|
model_inputs["chosen_labels"].append(chosen_labels) |
|
|
model_inputs["rejected_input_ids"].append(rejected_input_ids) |
|
|
model_inputs["rejected_attention_mask"].append([1] * len(rejected_input_ids)) |
|
|
model_inputs["rejected_labels"].append(rejected_labels) |
|
|
if processor is not None: |
|
|
model_inputs["pixel_values"].append(get_pixel_values(examples["images"][i], processor)) |
|
|
if hasattr(processor, "image_seq_length"): |
|
|
model_inputs["chosen_token_type_ids"].append( |
|
|
get_paligemma_token_type_ids(len(chosen_input_ids), processor) |
|
|
) |
|
|
model_inputs["rejected_token_type_ids"].append( |
|
|
get_paligemma_token_type_ids(len(rejected_input_ids), processor) |
|
|
) |
|
|
|
|
|
return model_inputs |
|
|
|
|
|
|
|
|
def print_pairwise_dataset_example(example: Dict[str, List[int]], tokenizer: "PreTrainedTokenizer") -> None: |
|
|
valid_chosen_labels = list(filter(lambda x: x != IGNORE_INDEX, example["chosen_labels"])) |
|
|
valid_rejected_labels = list(filter(lambda x: x != IGNORE_INDEX, example["rejected_labels"])) |
|
|
print("chosen_input_ids:\n{}".format(example["chosen_input_ids"])) |
|
|
print("chosen_inputs:\n{}".format(tokenizer.decode(example["chosen_input_ids"], skip_special_tokens=False))) |
|
|
print("chosen_label_ids:\n{}".format(example["chosen_labels"])) |
|
|
print("chosen_labels:\n{}".format(tokenizer.decode(valid_chosen_labels, skip_special_tokens=False))) |
|
|
print("rejected_input_ids:\n{}".format(example["rejected_input_ids"])) |
|
|
print("rejected_inputs:\n{}".format(tokenizer.decode(example["rejected_input_ids"], skip_special_tokens=False))) |
|
|
print("rejected_label_ids:\n{}".format(example["rejected_labels"])) |
|
|
print("rejected_labels:\n{}".format(tokenizer.decode(valid_rejected_labels, skip_special_tokens=False))) |
|
|
|
