ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

support control eos, fix #6345

Browse Source 
Former-commit-id: cb0f8399356bf372f3b7963f2565c3d504be0923
This commit is contained in:
hiyouga
2024-12-17 10:42:05 +00:00
parent 6522467ddb
commit 19ebc0e7a2
5 changed files with 21 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
src/llamafactory/train/sft/trainer.py
View File

@@ -141,7 +141,9 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
padded_tensor[:, -src_tensor.shape[-1] :] = src_tensor # adopt left-padding
return padded_tensor.contiguous() # in contiguous memory
def save_predictions(self, dataset: "Dataset", predict_results: "PredictionOutput") -> None:
def save_predictions(
self, dataset: "Dataset", predict_results: "PredictionOutput", gen_kwargs: Dict[str, Any]
) -> None:
r"""
Saves model predictions to `output_dir`.
@@ -168,8 +170,12 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
preds[i] = np.concatenate((preds[i][pad_len[0] :], preds[i][: pad_len[0]]), axis=-1)
decoded_inputs = self.processing_class.batch_decode(dataset["input_ids"], skip_special_tokens=False)
decoded_preds = self.processing_class.batch_decode(preds, skip_special_tokens=True)
decoded_labels = self.processing_class.batch_decode(labels, skip_special_tokens=True)
decoded_preds = self.processing_class.batch_decode(
preds, skip_special_tokens=gen_kwargs["skip_special_tokens"]
)
decoded_labels = self.processing_class.batch_decode(
labels, skip_special_tokens=gen_kwargs["skip_special_tokens"]
)
with open(output_prediction_file, "w", encoding="utf-8") as f:
for text, pred, label in zip(decoded_inputs, decoded_preds, decoded_labels):

2
src/llamafactory/train/sft/workflow.py
View File

@@ -130,7 +130,7 @@ def run_sft(
predict_results.metrics.pop("predict_loss", None)
trainer.log_metrics("predict", predict_results.metrics)
trainer.save_metrics("predict", predict_results.metrics)
trainer.save_predictions(dataset_module["eval_dataset"], predict_results)
trainer.save_predictions(dataset_module["eval_dataset"], predict_results, gen_kwargs)
# Create model card
create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)