423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-12-15 03:10:35 +08:00
Code Issues Packages Projects Releases Wiki Activity

add kto

Browse Source 
Former-commit-id: db1d5a4f51
This commit is contained in:
enji.zhou
2024-05-17 13:09:17 +08:00
parent 1bbbcb5895
commit 03956053b8
14 changed files with 5923 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
src/llamafactory/hparams/finetuning_args.py
View File

@@ -133,6 +133,22 @@ class RLHFArguments:
default=0.0,
metadata={"help": "The supervised fine-tuning loss coefficient in DPO training."},
)
kto_beta: float = field(
default=0.1,
metadata={"help": "The beta parameter for the KTO loss."},
)
kto_ftx: float = field(
default=0.0,
metadata={"help": "The supervised fine-tuning loss coefficient in KTO training."},
)
kto_desirable_weight: float = field(
default=1.0,
metadata={"help": "The desirable weight for the KTO loss."},
)
kto_undesirable_weight: float = field(
default=1.0,
metadata={"help": "The undesirable weight for the KTO loss."},
)
orpo_beta: float = field(
default=0.1,
metadata={"help": "The beta (lambda) parameter in ORPO loss representing the weight of the SFT loss."},
@@ -291,7 +307,7 @@ class FinetuningArguments(FreezeArguments, LoraArguments, RLHFArguments, GaloreA
default=False,
metadata={"help": "Whether or not to train model in purely bf16 precision (without AMP)."},
)
stage: Literal["pt", "sft", "rm", "ppo", "dpo", "orpo"] = field(
stage: Literal["pt", "sft", "rm", "ppo", "dpo", "orpo", "kto"] = field(
default="sft",
metadata={"help": "Which stage will be performed in training."},
)