ctopd / kdflow /arguments /data_args.py
codemaivanngu's picture
chore(upstream): import Cross-Tokenizer-OPD source snapshot
0ca529a unverified
Raw History Blame Contribute Delete
3.36 kB
from dataclasses import dataclass, field
from typing import Optional
@dataclass
class DataArguments:
""" Arguments for dataset."""
train_dataset_path: str = field(
default=None,
metadata={"help": "Training dataset name or path."}
)
train_dataset_probs: str = field(
default=None,
metadata={"help": "Sampling probs for multiple datasets."}
)
train_split: str = field(
default="train",
metadata={"help": "Train split in dataset."}
)
eval_dataset_path: str = field(
default=None,
metadata={"help": "Evaluation dataset name or path."}
)
eval_split: str = field(
default="eval",
metadata={"help": "Eval split in dataset."}
)
custom_eval_fn: str = field(
default=None,
metadata={
"help": "Optional Python file defining eval_fn(predictions, labels), "
"which must return a metric dict (only supported in on-policy KD)."
}
)
eval_math_dataset_path: str = field(
default=None,
metadata={
"help": "Optional math evaluation JSONL path. Each row must contain "
"'problem' and 'answer' fields (only supported in on-policy KD)."
}
)
eval_math_n_samples: int = field(
default=1,
metadata={"help": "Number of generated answers per math problem for pass@k evaluation."}
)
eval_math_max_concurrent: int = field(
default=256,
metadata={"help": "Maximum number of concurrent math evaluation requests."}
)
eval_math_max_tokens: int = field(
default=8192,
metadata={"help": "Maximum number of new tokens generated for math evaluation."}
)
eval_math_timeout: int = field(
default=1800,
metadata={"help": "Total timeout in seconds for math evaluation generation."}
)
input_key: str = field(
default="messages",
metadata={"help": "JSON dataset key."}
)
image_key: str = field(
default=None,
metadata={"help": "Image key in dataset."}
)
output_key: str = field(
default=None,
metadata={"help": "Output key in dataset."}
)
teacher_input_key: str = field(
default=None,
metadata={"help": "Input key for teacher prompt. If None, use the same input_key as student. "
"Used in self-distillation where teacher receives richer information."}
)
label_key: str = field(
default=None,
metadata={"help": "Label key in dataset."}
)
teacher_routing_key: str = field(
default="teacher_routing_key",
metadata={"help": "Route key for teacher model in multi-teacher distillation."}
)
input_template: str = field(
default=None,
)
apply_chat_template: bool = field(
default=True,
metadata={"help": "Use HF tokenizer chat template."}
)
max_len: int = field(
default=4096
)
max_samples: int = field(
default=1e8
)
packing_samples: bool = field(
default=False,
metadata={"help": "Packing sequences during training."}
)
prompt_max_len: int = field(
default=2048,
metadata={"help": "Max prompt length."}
)
preprocess_num_workers: int = field(
default=1,
)