Compute metrics for generation tasks in SFTTrainer #862
Description
Hi, I want to include a custom generation based compute_metrics e.g., BLEU, to the SFTTrainer. However, I have difficulties because:
- The input,
eval_preds, into compute_metrics contains a.predictionsattribute but its value are logits rather than generations. There don't seem to be an attribute for generations. - In
Seq2seqTrainingArguments, there is apredict_with_generateargument for the purpose of generation based metrics. But this does not work on theSFTTrainer. So I am not able to compute the metrics that I want.
Can anyone advise how I can get the generations into the compute_metrics function, please? The trl version I am using is trl==0.5.0
Here are my TrainingArgs and SFTTrainer.
training_args = Seq2SeqTrainingArguments(
output_dir=output_dir,
eval_steps=eval_steps,
evaluation_strategy=evaluation_strategy,
gradient_accumulation_steps=gradient_accumulation_steps,
learning_rate=lr,
max_steps=max_steps,
num_train_epochs=num_train_epochs,
per_device_train_batch_size=batch_size,
per_device_eval_batch_size=batch_size,
save_steps=save_steps,
save_strategy=save_strategy,
warmup_steps=num_warmup_steps,
weight_decay=weight_decay,
save_total_limit=save_total_limit,
load_best_model_at_end=True,
metric_for_best_model=metric_for_best_model,
greater_is_better=metric_greater_is_better,
generation_config=generation_config,
predict_with_generate=True,
generation_max_length=10,
)
trainer = SFTTrainer(
model,
args=training_args,
train_dataset=dataset_train,
eval_dataset=dataset_eval,
formatting_func=formatting_prompts_func,
data_collator=collator,
compute_metrics=compute_metrics,
)
Attributes of eval_preds are:
>> dir(eval_preds)
['__class__',
'__delattr__',
'__dict__',
'__dir__',
'__doc__',
'__eq__',
'__format__',
'__ge__',
'__getattribute__',
'__getitem__',
'__gt__',
'__hash__',
'__init__',
'__init_subclass__',
'__iter__',
'__le__',
'__lt__',
'__module__',
'__ne__',
'__new__',
'__reduce__',
'__reduce_ex__',
'__repr__',
'__setattr__',
'__sizeof__',
'__str__',
'__subclasshook__',
'__weakref__',
'inputs',
'label_ids',
'predictions']
Thank you in advance!