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

Merge pull request #225 from unslothai/fix/fix-response-on-completion-truncation

Browse source 
fix: error on >30% sample drop after `train_on_responses_only` instead of silent DataLoader crash
This commit is contained in:
Roland Tannous 2026-02-23 16:28:32 +04:00 • committed by GitHub
commit 2e2aa54ad2
1 changed files with 36 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

37
studio/backend/core/training/trainer.py
View file

@ -874,6 +874,41 @@ class UnslothTrainer:
num_proc=config_args.get("dataset_num_proc", max(1, os.cpu_count() // 4)),
)
print("Train on responses only configured successfully\n")
# ── Safety net: check if all samples were filtered out ──
# Unsloth's train_on_responses_only masks non-response
# tokens with -100. If max_seq_length is too short and the
# response portion gets truncated away, EVERY sample ends
# up with all labels == -100 and Unsloth removes them,
# leaving 0 usable training samples.
filtered_len = len(self.trainer.train_dataset)
original_len = len(dataset["dataset"])
dropped = original_len - filtered_len
drop_pct = round(100 * dropped / original_len, 1) if original_len > 0 else 0
if filtered_len == 0 or drop_pct > 30:
max_seq = training_args.get('max_seq_length', 2048)
error_msg = (
f"{dropped}/{original_len} samples ({drop_pct}%) "
f"were dropped after applying 'train on responses "
f"only' — only {filtered_len} remain. This usually "
f"means max_seq_length ({max_seq}) is too short "
f"and the response portion is being truncated "
f"away. Try increasing max_seq_length (e.g. 8192) "
f"or disabling 'Train on completions'."
)
logger.error(error_msg)
self._update_progress(error=error_msg, is_training=False)
return
if dropped > 0:
print(
f"⚠️ {dropped}/{original_len} samples "
f"({drop_pct}%) were dropped (all labels "
f"masked). {filtered_len} samples remain.\n"
)
print(f"Post-filter dataset size: {filtered_len} samples\n")
except Exception as e:
logger.warning(f"Failed to apply train on responses only: {e}")
train_on_responses_enabled = False
@ -955,7 +990,7 @@ class UnslothTrainer:
progress_callback = ProgressCallback(self)
self.trainer.add_callback(progress_callback)
num_samples = len(dataset["dataset"])
num_samples = len(self.trainer.train_dataset)
batch_size = training_args.get('batch_size', 2)
grad_accum = training_args.get('gradient_accumulation_steps', 4)
num_epochs = training_args.get('num_epochs', 3)