Skip to content
New issue

Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.

#

By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? # to your account

Fix: --disable_group_texts 1 keep short samples #649

Merged
merged 1 commit into from
Sep 22, 2023
Merged

Fix: --disable_group_texts 1 keep short samples #649

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 17 additions & 3 deletions src/lmflow/models/hf_decoder_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -493,6 +493,16 @@ def tokenize(self, dataset, add_special_tokens=True, *args, **kwargs):
# logger loading before tokenize_function
tok_logger = transformers.utils.logging.get_logger("transformers.tokenization_utils_base")

data_args = raw_datasets.get_data_args()

# Whether to truncate long sequences to fit into max_length
use_truncation = False
if model_args.use_lora or data_args.disable_group_texts:
use_truncation = True

# Whether to pad short sequences to max_length
padding = "max_length" if data_args.disable_group_texts else False

def tokenize_function(examples):
num_example = len(examples[column_names[0]])
token_dict = {
Expand All @@ -505,7 +515,8 @@ def tokenize_function(examples):
encoding = self.tokenizer(
examples[column_name],
add_special_tokens=add_special_tokens,
truncation=True if model_args.use_lora else None,
truncation=use_truncation,
padding=padding,
)

if column_name in label_columns:
Expand Down Expand Up @@ -533,11 +544,14 @@ def tokenize_function(examples):
)
return token_dict

data_args = raw_datasets.get_data_args()
if not data_args.streaming:
fingerprint = raw_datasets.get_fingerprint()
new_fingerprint = hashlib.md5(
(fingerprint + str(self.tokenizer)).encode("utf-8")
(
fingerprint
+ str(self.tokenizer)
+ f'###disable_group_texts={data_args.disable_group_texts}'
).encode("utf-8")
).hexdigest()

tokenized_datasets = raw_datasets.map(
Expand Down
24 changes: 16 additions & 8 deletions src/lmflow/pipeline/finetuner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -230,24 +230,32 @@ def tune(self,
else:
with finetuner_args.main_process_first(desc="dataset map tokenization"):
tokenized_dataset = model.tokenize(dataset)
lm_dataset = self.group_text(
tokenized_dataset,
model_max_length=model.get_max_length(),
)
if data_args.disable_group_texts:
lm_dataset = tokenized_dataset
else:
lm_dataset = self.group_text(
tokenized_dataset,
model_max_length=model.get_max_length(),
)

train_dataset = lm_dataset.get_backend_dataset()
logger.info(f"Number of train samples: {len(train_dataset)}")

if finetuner_args.do_eval:
eval_dataset_args = deepcopy(data_args)
eval_dataset_args.dataset_path = finetuner_args.eval_dataset_path
eval_dataset = Dataset(eval_dataset_args)
with finetuner_args.main_process_first(desc="dataset map tokenization"):
tokenized_dataset = model.tokenize(eval_dataset)
lm_dataset = self.group_text(
tokenized_dataset,
model_max_length=model.get_max_length(),
)
if data_args.disable_group_texts:
lm_dataset = tokenized_dataset
else:
lm_dataset = self.group_text(
tokenized_dataset,
model_max_length=model.get_max_length(),
)
eval_dataset = lm_dataset.get_backend_dataset()
logger.info(f"Number of eval samples: {len(train_dataset)}")


def preprocess_logits_for_metrics(logits, labels):
Expand Down