ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

skip eval split and HF split detection when eval_steps is disabled

Browse source
This commit is contained in:
Roland Tannous 2026-02-24 09:26:54 +00:00
commit 2be2933846
2 changed files with 24 additions and 22 deletions
Download patch file Download diff file Expand all files Collapse all files

41
studio/backend/core/training/trainer.py
View file

@ -343,7 +343,8 @@ class UnslothTrainer:
custom_format_mapping: dict = None,
subset: str = None,
train_split: str = "train",
eval_split: str = None) -> Optional[tuple]:
eval_split: str = None,
eval_steps: float = 0.00) -> Optional[tuple]:
"""
Load and prepare dataset for training.
@ -358,6 +359,7 @@ class UnslothTrainer:
dataset = None
eval_dataset = None
has_separate_eval_source = False # True if eval comes from a separate HF split
eval_enabled = eval_steps is not None and eval_steps > 0
if local_datasets:
# Load local datasets
@ -410,23 +412,26 @@ class UnslothTrainer:
print(f"Loaded dataset from Hugging Face: {dataset_source}\n")
# Resolve eval split from a separate HF split (explicit or auto-detected)
if eval_split:
# Explicit eval split provided - load it directly
print(f"Loading explicit eval split: '{eval_split}'\n")
eval_load_kwargs = {"path": dataset_source, "split": eval_split}
if subset:
eval_load_kwargs["name"] = subset
eval_dataset = load_dataset(**eval_load_kwargs)
has_separate_eval_source = True
print(f"Loaded eval split '{eval_split}' with {len(eval_dataset)} rows\n")
else:
# Auto-detect eval split from HF (returns a separate dataset, or None)
eval_dataset = self._auto_detect_eval_split_from_hf(
dataset_source=dataset_source,
subset=subset,
)
if eval_dataset is not None:
if eval_enabled:
if eval_split:
# Explicit eval split provided - load it directly
print(f"Loading explicit eval split: '{eval_split}'\n")
eval_load_kwargs = {"path": dataset_source, "split": eval_split}
if subset:
eval_load_kwargs["name"] = subset
eval_dataset = load_dataset(**eval_load_kwargs)
has_separate_eval_source = True
print(f"Loaded eval split '{eval_split}' with {len(eval_dataset)} rows\n")
else:
# Auto-detect eval split from HF (returns a separate dataset, or None)
eval_dataset = self._auto_detect_eval_split_from_hf(
dataset_source=dataset_source,
subset=subset,
)
if eval_dataset is not None:
has_separate_eval_source = True
else:
print("Eval disabled (eval_steps <= 0), skipping eval split detection\n")
if dataset is None:
raise ValueError("No dataset provided")
@ -472,7 +477,7 @@ class UnslothTrainer:
)
eval_dataset = eval_info["dataset"]
print(f"Eval dataset formatted successfully\n")
elif not has_separate_eval_source:
elif eval_enabled and not has_separate_eval_source:
# No separate eval source — split the already-formatted dataset
formatted_dataset = dataset_info["dataset"]
split_result = self._resolve_eval_split_from_dataset(formatted_dataset)

5
studio/backend/core/training/training.py
View file

@ -223,6 +223,7 @@ class TrainingBackend:
subset=subset,
train_split=train_split,
eval_split=eval_split,
eval_steps=eval_steps,
)
# Unpack: load_and_format_dataset returns (dataset, eval_dataset)
@ -232,10 +233,6 @@ class TrainingBackend:
dataset = dataset_result
eval_dataset = None
# If user set eval_steps to 0, disable evaluation entirely
if eval_steps is not None and float(eval_steps) <= 0:
eval_dataset = None
# Track whether eval is enabled for status reporting
self.eval_enabled = eval_dataset is not None