fix: padding 默认值改为 0,避免 loss_mask 误计 padding 位置
This commit is contained in:
@@ -83,7 +83,7 @@ def cache_jsonl(
|
|||||||
processor: BaseProcessor,
|
processor: BaseProcessor,
|
||||||
*,
|
*,
|
||||||
pack_size: int = -1,
|
pack_size: int = -1,
|
||||||
pad_value: int = 1,
|
pad_value: int = 0,
|
||||||
) -> List[str]:
|
) -> List[str]:
|
||||||
"""Tokenize JSONL files and pack them into HDF5 storage.
|
"""Tokenize JSONL files and pack them into HDF5 storage.
|
||||||
|
|
||||||
|
|||||||
+1
-1
@@ -51,7 +51,7 @@ def main():
|
|||||||
help="Pack size, <=0 to disable (default: -1)",
|
help="Pack size, <=0 to disable (default: -1)",
|
||||||
)
|
)
|
||||||
parser.add_argument(
|
parser.add_argument(
|
||||||
"--pad-value", type=int, default=1, help="Padding value (default: 1)"
|
"--pad-value", type=int, default=0, help="Padding value (default: 0)"
|
||||||
)
|
)
|
||||||
parser.add_argument(
|
parser.add_argument(
|
||||||
"--log-level",
|
"--log-level",
|
||||||
|
|||||||
Reference in New Issue
Block a user