s0duku's picture
Add files using upload-large-folder tool
4b10a0d verified
Raw History Blame Contribute Delete
2.59 kB
"""Immutable configuration and output models for repository sample building."""
from __future__ import annotations
from dataclasses import dataclass, field
from typing import Any, Mapping
BEGIN_REPO_NAME = "<|begin▁of▁repo▁name|>"
END_REPO_NAME = "<|end▁of▁repo▁name|>"
BEGIN_FILE_NAME = "<|begin▁of▁file▁name|>"
END_FILE_NAME = "<|end▁of▁file▁name|>"
BEGIN_FILE = "<|begin▁of▁file|>"
END_FILE = "<|end▁of▁file|>"
SPECIAL_TOKENS = (
BEGIN_REPO_NAME,
END_REPO_NAME,
BEGIN_FILE_NAME,
END_FILE_NAME,
BEGIN_FILE,
END_FILE,
)
@dataclass(frozen=True, slots=True)
class BuildConfig:
"""Controls deterministic repository sample construction.
``max_tokens`` is enforced with the supplied tokenizer, or the pinned
DeepSeek V4 Flash Base tokenizer by default.
"""
max_tokens: int = 8_192
min_tokens: int = 1
max_file_tokens: int | None = None
include_weak_edges: bool = False
metadata_only_vendor: bool = True
metadata_only_generated: bool = True
target_context: bool = False
max_targets_per_repo: int = 32
max_dependency_depth: int = 3
utf8_bytes_per_token: float = 3.0
def __post_init__(self) -> None:
if self.max_tokens < 64:
raise ValueError("max_tokens must be at least 64")
if self.min_tokens < 0 or self.min_tokens > self.max_tokens:
raise ValueError("min_tokens must be between 0 and max_tokens")
if self.max_file_tokens is not None and self.max_file_tokens < 1:
raise ValueError("max_file_tokens must be positive")
if self.max_targets_per_repo < 0:
raise ValueError("max_targets_per_repo must be non-negative")
if self.max_dependency_depth < 0:
raise ValueError("max_dependency_depth must be non-negative")
if self.utf8_bytes_per_token <= 0:
raise ValueError("utf8_bytes_per_token must be positive")
@dataclass(frozen=True, slots=True)
class SampleFile:
"""Provenance for one repository file represented in a sample."""
path: str
relation: str
content_policy: str
role: str = "unknown"
language: str | None = None
bytes: int = 0
part: int = 1
parts: int = 1
@dataclass(frozen=True, slots=True)
class Sample:
"""One rendered training document and its complete provenance."""
sample_id: str
repo_id: str
revision_id: str
kind: str
text: str
files: tuple[SampleFile, ...]
token_count: int
metadata: Mapping[str, Any] = field(default_factory=dict)