tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

additional logging to get maximum token length of a sequence in the dataset (#1066) [skip ci]

Browse Source 
* additional logging to get maximum token length of a sequence in the dataset

* fix ordering to properly determine the max_len of tokens before dropping anything longer
This commit is contained in:
Wing Lian
2024-01-10 00:49:31 -05:00
committed by GitHub
parent 0ce1a6594e
commit 2f2582e6ed
1 changed files with 10 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
src/axolotl/utils/trainer.py
View File

@@ -109,12 +109,6 @@ def disable_datasets_caching():
def process_datasets_for_packing(cfg, train_dataset, eval_dataset, tokenizer): def process_datasets_for_packing(cfg, train_dataset, eval_dataset, tokenizer):
drop_long = partial(drop_long_seq, sequence_len=cfg.sequence_len) drop_long = partial(drop_long_seq, sequence_len=cfg.sequence_len)
with zero_first(is_main_process()): with zero_first(is_main_process()):
train_dataset = train_dataset.filter(drop_long, num_proc=cfg.dataset_processes)
if eval_dataset:
eval_dataset = eval_dataset.filter(
drop_long, num_proc=cfg.dataset_processes
)
if cfg.group_by_length: if cfg.group_by_length:
train_dataset = train_dataset.map( train_dataset = train_dataset.map(
add_length, num_proc=cfg.dataset_processes add_length, num_proc=cfg.dataset_processes
@@ -130,6 +124,16 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset, tokenizer):
add_position_ids, num_proc=cfg.dataset_processes add_position_ids, num_proc=cfg.dataset_processes
) )
if cfg.group_by_length or cfg.sample_packing:
max_input_len = np.max(get_dataset_lengths(train_dataset))
LOG.debug(f"max_input_len: {max_input_len}", main_process_only=True)
train_dataset = train_dataset.filter(drop_long, num_proc=cfg.dataset_processes)
if eval_dataset:
eval_dataset = eval_dataset.filter(
drop_long, num_proc=cfg.dataset_processes
)
# Phi doesn't want the attention_mask feature when training # Phi doesn't want the attention_mask feature when training
if ( if (
"CodeGenTokenizer" in tokenizer.__class__.__name__ "CodeGenTokenizer" in tokenizer.__class__.__name__