`bridge.data.datasets.preference`#

Module Contents#

Scoring-run inputs a training run must reproduce; written to and checked against scoring_metadata.json.

`_open_path`
`ref_logprobs_from_rows`	Validate offline-scorer output rows and key them by `pair_id`.
`_artifact_path`
`write_ref_logprobs`	Write the ref-logprob rows as JSONL to `artifact_dir` (local or msc://).
`load_ref_logprobs`	Load and validate a ref-scorer artifact from `path` (local or msc://).
`write_scoring_metadata`	Write `scoring_metadata.json` to `artifact_dir` (local or msc://).
`read_scoring_metadata`	Load the artifact’s `scoring_metadata.json` (local or msc://).
`validate_scoring_metadata`	Raise if the artifact was scored under different inputs than `expected`; return its metadata.
`pair_token_lengths`	Per-pair padded row length (max of the two sides), one full pass over `dataset`.
`pack_pairs_by_token_budget`	Deterministic longest-first batches under `budget_tokens`; an over-budget pair gets a singleton.
`preference_collate_fn`	Collate pair records into an interleaved row batch (chosen@even, rejected@odd).
`build_preference_data_loader`	Build a DPO loader from row-denominated sizes (one pair == two rows), halved here for the stock loader.

bridge.data.datasets.preference.SCORING_METADATA_FILENAME#: ‘scoring_metadata.json’

bridge.data.datasets.preference.REF_LOGPROBS_FILENAME#: ‘ref_logprobs.jsonl’

class bridge.data.datasets.preference.ScoringFingerprint#

Scoring-run inputs a training run must reproduce; written to and checked against scoring_metadata.json.

tokenization_version: int#

None

Bumped whenever the pair token streams change for the same inputs, so older artifacts fail the check.

bridge.data.datasets.preference.REF_LOGPROB_COLUMNS: tuple[str, ...]#: (‘ref_chosen_logprob_sum’, ‘ref_chosen_num_tokens’, ‘ref_rejected_logprob_sum’, ‘ref_rejected_num_to…

bridge.data.datasets.preference.ref_logprobs_from_rows( rows: collections.abc.Iterable[collections.abc.Mapping[str, Any]], expected_num_pairs: int, ) → dict[int, dict[str, Any]]#: Validate offline-scorer output rows and key them by pair_id.

bridge.data.datasets.preference._artifact_path(artifact_dir: str, filename: str) → str#

bridge.data.datasets.preference.write_ref_logprobs( records: collections.abc.Sequence[collections.abc.Mapping[str, Any]], artifact_dir: str, ) → None#: Write the ref-logprob rows as JSONL to artifact_dir (local or msc://).

bridge.data.datasets.preference.load_ref_logprobs( path: str, expected_num_pairs: int, ) → dict[int, dict[str, Any]]#: Load and validate a ref-scorer artifact from path (local or msc://).

bridge.data.datasets.preference.write_scoring_metadata( metadata: collections.abc.Mapping[str, Any], artifact_dir: str, ) → None#: Write scoring_metadata.json to artifact_dir (local or msc://).

bridge.data.datasets.preference.read_scoring_metadata(artifact_dir: str) → dict[str, Any]#: Load the artifact’s scoring_metadata.json (local or msc://).

bridge.data.datasets.preference.validate_scoring_metadata( artifact_dir: str, expected: bridge.data.datasets.preference.ScoringFingerprint, ) → dict[str, Any]#: Raise if the artifact was scored under different inputs than expected; return its metadata.

bridge.data.datasets.preference.pair_token_lengths(dataset: torch.utils.data.Dataset) → list[int]#: Per-pair padded row length (max of the two sides), one full pass over dataset.

bridge.data.datasets.preference.pack_pairs_by_token_budget( lengths: collections.abc.Sequence[int], *, budget_tokens: int, max_pairs_per_batch: int = 64, ) → list[list[int]]#: Deterministic longest-first batches under budget_tokens; an over-budget pair gets a singleton.

bridge.data.datasets.preference.preference_collate_fn( batch: list[collections.abc.Mapping[str, Any]], pad_token_id: int, pad_seq_length_to_mult: int = 1, require_ref_logprobs: bool = True, ) → dict[str, Any]#: Collate pair records into an interleaved row batch (chosen@even, rejected@odd).

bridge.data.datasets.preference.build_preference_data_loader( dataset: torch.utils.data.Dataset, micro_batch_size: int, global_batch_size: int, data_parallel_rank: int, data_parallel_size: int, consumed_samples: int, num_workers: int = 0, pin_memory: bool = True, worker_init_fn: collections.abc.Callable[[int], None] | None = None, shuffle: bool = True, seed: int | None = None, ) → torch.utils.data.DataLoader#: Build a DPO loader from row-denominated sizes (one pair == two rows), halved here for the stock loader.