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

[misc] fix packing and eval plot (#7623)

Browse Source
This commit is contained in:
hoshi-hiyouga
2025-04-07 18:20:57 +08:00
committed by GitHub
parent 5115dc8c7f
commit c3c0efbaa0
70 changed files with 288 additions and 194 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/llamafactory/train/dpo/workflow.py
View File

@@ -91,7 +91,13 @@ def run_dpo(
trainer.save_metrics("train", train_result.metrics)
trainer.save_state()
if trainer.is_world_process_zero() and finetuning_args.plot_loss:
plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "rewards/accuracies"])
keys = ["loss", "rewards/accuracies"]
if isinstance(dataset_module["eval_dataset"], dict):
keys += [f"eval_{key}_loss" for key in dataset_module["eval_dataset"].keys()]
else:
keys += ["eval_loss"]
plot_loss(training_args.output_dir, keys=keys)
# Evaluation
if training_args.do_eval:

8
src/llamafactory/train/kto/workflow.py
View File

@@ -82,7 +82,13 @@ def run_kto(
trainer.save_metrics("train", train_result.metrics)
trainer.save_state()
if trainer.is_world_process_zero() and finetuning_args.plot_loss:
plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "rewards/chosen"])
keys = ["loss", "rewards/chosen"]
if isinstance(dataset_module["eval_dataset"], dict):
keys += [f"eval_{key}_loss" for key in dataset_module["eval_dataset"].keys()]
else:
keys += ["eval_loss"]
plot_loss(training_args.output_dir, keys=keys)
# Evaluation
if training_args.do_eval:

8
src/llamafactory/train/pt/workflow.py
View File

@@ -66,7 +66,13 @@ def run_pt(
trainer.save_metrics("train", train_result.metrics)
trainer.save_state()
if trainer.is_world_process_zero() and finetuning_args.plot_loss:
plot_loss(training_args.output_dir, keys=["loss", "eval_loss"])
keys = ["loss"]
if isinstance(dataset_module["eval_dataset"], dict):
keys += [f"eval_{key}_loss" for key in dataset_module["eval_dataset"].keys()]
else:
keys += ["eval_loss"]
plot_loss(training_args.output_dir, keys=keys)
# Evaluation
if training_args.do_eval:

10
src/llamafactory/train/rm/workflow.py
View File

@@ -74,7 +74,15 @@ def run_rm(
trainer.save_metrics("train", train_result.metrics)
trainer.save_state()
if trainer.is_world_process_zero() and finetuning_args.plot_loss:
plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "eval_accuracy"])
keys = ["loss"]
if isinstance(dataset_module["eval_dataset"], dict):
keys += sum(
[[f"eval_{key}_loss", f"eval_{key}_accuracy"] for key in dataset_module["eval_dataset"].keys()], []
)
else:
keys += ["eval_loss", "eval_accuracy"]
plot_loss(training_args.output_dir, keys=keys)
# Evaluation
if training_args.do_eval:

10
src/llamafactory/train/sft/workflow.py
View File

@@ -110,7 +110,15 @@ def run_sft(
trainer.save_metrics("train", train_result.metrics)
trainer.save_state()
if trainer.is_world_process_zero() and finetuning_args.plot_loss:
plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "eval_accuracy"])
keys = ["loss"]
if isinstance(dataset_module["eval_dataset"], dict):
keys += sum(
[[f"eval_{key}_loss", f"eval_{key}_accuracy"] for key in dataset_module["eval_dataset"].keys()], []
)
else:
keys += ["eval_loss", "eval_accuracy"]
plot_loss(training_args.output_dir, keys=keys)
if training_args.predict_with_generate:
tokenizer.padding_side = "left" # use left-padding in generation