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

debug: decode first sample after train_on_completions masking

Browse source
This commit is contained in:
Roland Tannous 2026-03-10 14:08:14 +00:00
commit 7f1fd28acd
1 changed files with 10 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

10
studio/backend/core/training/trainer.py
View file

@ -2690,6 +2690,16 @@ class UnslothTrainer:
)
print(f"Post-filter dataset size: {filtered_len} samples\n")
# [DEBUG] Decode first sample AFTER train_on_completions applied
try:
_row = self.trainer.train_dataset[0]
_space = self.tokenizer(" ", add_special_tokens=False).input_ids[0]
print("[DEBUG] === After train_on_completions ===", flush=True)
print(f"[DEBUG] input_ids decoded:\n{self.tokenizer.decode(_row['input_ids'])}\n", flush=True)
print(f"[DEBUG] labels decoded (-100 → space):\n{self.tokenizer.decode([_space if x == -100 else x for x in _row['labels']])}\n", flush=True)
except Exception as _dbg_e:
print(f"[DEBUG] Could not decode post-completions sample: {_dbg_e}", flush=True)
except Exception as e:
logger.warning(f"Failed to apply train on responses only: {e}")
train_on_responses_enabled = False