feat: add record-mode to Store for DPO/GRPO
- Store gains fetch_record/num_records alongside stream fetch/__len__ - save_bin/load_bin support per-record offsets via record_keys param - H5Store/MmapStore/JsonlStore all support dual stream+record access - DPODataset/GRPODataset use fetch_record, no cross-record concat - dpo_collate_fn + collate_fn wired through TrainConfig - fixes attention context leakage in DPO from windowed concatenation
This commit is contained in:
@@ -1,9 +1,10 @@
|
||||
from astrai.dataset.dataset import (
|
||||
BaseDataset,
|
||||
DatasetFactory,
|
||||
dpo_collate_fn,
|
||||
grpo_collate_fn,
|
||||
)
|
||||
from astrai.dataset.sampler import ResumableDistributedSampler
|
||||
from astrai.dataset.sampler import RDSampler
|
||||
from astrai.dataset.storage import (
|
||||
H5Store,
|
||||
JsonlStore,
|
||||
@@ -22,6 +23,7 @@ from astrai.serialization import (
|
||||
__all__ = [
|
||||
"BaseDataset",
|
||||
"DatasetFactory",
|
||||
"dpo_collate_fn",
|
||||
"grpo_collate_fn",
|
||||
"Store",
|
||||
"StoreFactory",
|
||||
@@ -33,5 +35,5 @@ __all__ = [
|
||||
"load_h5",
|
||||
"save_bin",
|
||||
"load_bin",
|
||||
"ResumableDistributedSampler",
|
||||
"RDSampler",
|
||||
]
|
||||
|
||||
Reference in New Issue
Block a user