Skip to content

feat: add is_vlm param to safe_conversations_generator for multimodal data support #545

New issue

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

Sign up for GitHub

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

Already on GitHub? Sign in to your account

Open
sunny-infra wants to merge 5 commits into
base: main
Choose a base branch
from
Open

feat: add is_vlm param to safe_conversations_generator for multimodal data support #545

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
41ba68b
add the configs for qwen3-vl-8b-instruct model
sunny-infra Apr 23, 2026
b894fbc
feat: add is_vlm param to safe_conversations_generator for multimodal…
sunny-infra Apr 24, 2026
9defccf
feat: add is_vlm param to safe_conversations_generator for multimodal…
sunny-infra Apr 24, 2026
eadc83b
fix: fixed some formatting issues.
sunny-infra Apr 24, 2026
a91c698
fix: fixed some formatting issues v2
sunny-infra Apr 24, 2026
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
2 changes: 1 addition & 1 deletion scripts/prepare_hidden_states.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -618,7 +618,7 @@ def main():
print_with_rank("Loading/building dataset cache...")
dataset = Dataset.from_generator(
generator=safe_conversations_generator,
gen_kwargs={"file_path": args.data_path},
gen_kwargs={"file_path": args.data_path, "is_vlm": args.is_vlm},
cache_dir=os.path.join(
os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
"cache",
Expand Down
4 changes: 2 additions & 2 deletions scripts/train_eagle3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -456,7 +456,7 @@ def build_dataloaders(
cache_key = hashlib.md5(cache_params_string.encode()).hexdigest()
train_dataset = Dataset.from_generator(
generator=safe_conversations_generator,
gen_kwargs={"file_path": args.train_data_path},
gen_kwargs={"file_path": args.train_data_path, "is_vlm": args.is_vlm},
)
is_online = (
args.train_data_path is not None and args.train_hidden_states_path is None
Expand Down Expand Up @@ -507,7 +507,7 @@ def build_dataloaders(
if args.eval_data_path is not None:
eval_dataset = Dataset.from_generator(
generator=safe_conversations_generator,
gen_kwargs={"file_path": args.eval_data_path},
gen_kwargs={"file_path": args.eval_data_path, "is_vlm": args.is_vlm},
)
eval_eagle3_dataset = build_eagle3_dataset(
eval_dataset,
Expand Down
15 changes: 12 additions & 3 deletions specforge/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -328,12 +328,15 @@ def shard_optimizer_state_with_dtensor(bf16_optimizer, device_mesh):
)


def safe_conversations_generator(file_path):
def safe_conversations_generator(file_path, is_vlm=False):
"""
Generator that:
1. Extracts the 'conversations' field.
2. Preserves all original fields within each message.
3. [Key step] Converts all list/dict-type field values to strings to resolve mixed-type conflicts (e.g., for Arrow compatibility).
Args:
file_path: Path to the JSONL file.
is_vlm: If True, include 'image' field for vision-language models. Default False.
"""
with open(file_path, "r", encoding="utf-8") as f:
for i, line in enumerate(f):
Expand Down Expand Up @@ -376,8 +379,14 @@ def safe_conversations_generator(file_path):

cleaned_convs.append(new_msg)

# Build result with conversations
result = {"conversations": cleaned_convs}
if is_vlm:
image = row.get("image", "")
result = {
"conversations": cleaned_convs,
"image": image,
}
else:
result = {"conversations": cleaned_convs}

# Preserve 'tools' field if present
if "tools" in row:
Expand Down