together 1.5.13__py3-none-any.whl → 1.5.14__py3-none-any.whl

together/cli/api/finetune.py

@@ -142,6 +142,30 @@ def fine_tuning(ctx: click.Context) -> None:
     default=0.1,
     help="Beta parameter for DPO training (only used when '--training-method' is 'dpo')",
 )
+@click.option(
+    "--dpo-normalize-logratios-by-length",
+    type=bool,
+    default=False,
+    help=(
+        "Whether to normalize logratios by sample length "
+        "(only used when '--training-method' is 'dpo')"
+    ),
+)
+@click.option(
+    "--rpo-alpha",
+    type=float,
+    default=0.0,
+    help=(
+        "RPO alpha parameter of DPO training to include NLL in the loss "
+        "(only used when '--training-method' is 'dpo')"
+    ),
+)
+@click.option(
+    "--simpo-gamma",
+    type=float,
+    default=0.1,
+    help="SimPO gamma parameter (only used when '--training-method' is 'dpo')",
+)
 @click.option(
     "--suffix",
     "-s",
@@ -206,6 +230,9 @@ def create(
     train_on_inputs: bool | Literal["auto"],
     training_method: str,
     dpo_beta: float,
+    dpo_normalize_logratios_by_length: bool,
+    rpo_alpha: float,
+    simpo_gamma: float,
     from_checkpoint: str,
 ) -> None:
     """Start fine-tuning"""
@@ -239,6 +266,9 @@ def create(
         train_on_inputs=train_on_inputs,
         training_method=training_method,
         dpo_beta=dpo_beta,
+        dpo_normalize_logratios_by_length=dpo_normalize_logratios_by_length,
+        rpo_alpha=rpo_alpha,
+        simpo_gamma=simpo_gamma,
         from_checkpoint=from_checkpoint,
     )
 

together/resources/finetune.py

@@ -72,6 +72,9 @@ def create_finetune_request(
     train_on_inputs: bool | Literal["auto"] | None = None,
     training_method: str = "sft",
     dpo_beta: float | None = None,
+    dpo_normalize_logratios_by_length: bool = False,
+    rpo_alpha: float | None = None,
+    simpo_gamma: float | None = None,
     from_checkpoint: str | None = None,
 ) -> FinetuneRequest:
     if model is not None and from_checkpoint is not None:
@@ -182,6 +185,21 @@ def create_finetune_request(
 
     if dpo_beta is not None and training_method != "dpo":
         raise ValueError("dpo_beta is only supported for DPO training")
+    if dpo_normalize_logratios_by_length and training_method != "dpo":
+        raise ValueError(
+            "dpo_normalize_logratios_by_length=True is only supported for DPO training"
+        )
+    if rpo_alpha is not None:
+        if training_method != "dpo":
+            raise ValueError("rpo_alpha is only supported for DPO training")
+        if not rpo_alpha >= 0.0:
+            raise ValueError(f"rpo_alpha should be non-negative (got {rpo_alpha})")
+
+    if simpo_gamma is not None:
+        if training_method != "dpo":
+            raise ValueError("simpo_gamma is only supported for DPO training")
+        if not simpo_gamma >= 0.0:
+            raise ValueError(f"simpo_gamma should be non-negative (got {simpo_gamma})")
 
     lr_scheduler: FinetuneLRScheduler
     if lr_scheduler_type == "cosine":
@@ -204,7 +222,24 @@ def create_finetune_request(
     if training_method == "sft":
         training_method_cls = TrainingMethodSFT(train_on_inputs=train_on_inputs)
     elif training_method == "dpo":
-        training_method_cls = TrainingMethodDPO(dpo_beta=dpo_beta)
+        if simpo_gamma is not None and simpo_gamma > 0:
+            dpo_reference_free = True
+            dpo_normalize_logratios_by_length = True
+            rprint(
+                f"Parameter simpo_gamma was set to {simpo_gamma}. "
+                "SimPO training detected. Reference logits will not be used "
+                "and length normalization of log-probabilities will be enabled."
+            )
+        else:
+            dpo_reference_free = False
+
+        training_method_cls = TrainingMethodDPO(
+            dpo_beta=dpo_beta,
+            dpo_normalize_logratios_by_length=dpo_normalize_logratios_by_length,
+            dpo_reference_free=dpo_reference_free,
+            rpo_alpha=rpo_alpha,
+            simpo_gamma=simpo_gamma,
+        )
 
     finetune_request = FinetuneRequest(
         model=model,
@@ -302,6 +337,9 @@ class FineTuning:
         train_on_inputs: bool | Literal["auto"] | None = None,
         training_method: str = "sft",
         dpo_beta: float | None = None,
+        dpo_normalize_logratios_by_length: bool = False,
+        rpo_alpha: float | None = None,
+        simpo_gamma: float | None = None,
         from_checkpoint: str | None = None,
     ) -> FinetuneResponse:
         """
@@ -353,6 +391,9 @@ class FineTuning:
             training_method (str, optional): Training method. Defaults to "sft".
                 Supported methods: "sft", "dpo".
             dpo_beta (float, optional): DPO beta parameter. Defaults to None.
+            dpo_normalize_logratios_by_length (bool): Whether or not normalize logratios by sample length. Defaults to False,
+            rpo_alpha (float, optional): RPO alpha parameter of DPO training to include NLL in the loss. Defaults to None.
+            simpo_gamma: (float, optional): SimPO gamma parameter. Defaults to None.
             from_checkpoint (str, optional): The checkpoint identifier to continue training from a previous fine-tuning job.
                 The format: {$JOB_ID/$OUTPUT_MODEL_NAME}:{$STEP}.
                 The step value is optional, without it the final checkpoint will be used.
@@ -405,6 +446,9 @@ class FineTuning:
             train_on_inputs=train_on_inputs,
             training_method=training_method,
             dpo_beta=dpo_beta,
+            dpo_normalize_logratios_by_length=dpo_normalize_logratios_by_length,
+            rpo_alpha=rpo_alpha,
+            simpo_gamma=simpo_gamma,
             from_checkpoint=from_checkpoint,
         )
 
@@ -714,6 +758,9 @@ class AsyncFineTuning:
         train_on_inputs: bool | Literal["auto"] | None = None,
         training_method: str = "sft",
         dpo_beta: float | None = None,
+        dpo_normalize_logratios_by_length: bool = False,
+        rpo_alpha: float | None = None,
+        simpo_gamma: float | None = None,
         from_checkpoint: str | None = None,
     ) -> FinetuneResponse:
         """
@@ -765,6 +812,9 @@ class AsyncFineTuning:
             training_method (str, optional): Training method. Defaults to "sft".
                 Supported methods: "sft", "dpo".
             dpo_beta (float, optional): DPO beta parameter. Defaults to None.
+            dpo_normalize_logratios_by_length (bool): Whether or not normalize logratios by sample length. Defaults to False,
+            rpo_alpha (float, optional): RPO alpha parameter of DPO training to include NLL in the loss. Defaults to None.
+            simpo_gamma: (float, optional): SimPO gamma parameter. Defaults to None.
             from_checkpoint (str, optional): The checkpoint identifier to continue training from a previous fine-tuning job.
                 The format: {$JOB_ID/$OUTPUT_MODEL_NAME}:{$STEP}.
                 The step value is optional, without it the final checkpoint will be used.
@@ -817,6 +867,9 @@ class AsyncFineTuning:
             train_on_inputs=train_on_inputs,
             training_method=training_method,
             dpo_beta=dpo_beta,
+            dpo_normalize_logratios_by_length=dpo_normalize_logratios_by_length,
+            rpo_alpha=rpo_alpha,
+            simpo_gamma=simpo_gamma,
             from_checkpoint=from_checkpoint,
         )
 

together/types/finetune.py

@@ -159,6 +159,10 @@ class TrainingMethodDPO(TrainingMethod):
 
     method: Literal["dpo"] = "dpo"
     dpo_beta: float | None = None
+    dpo_normalize_logratios_by_length: bool = False
+    dpo_reference_free: bool = False
+    rpo_alpha: float | None = None
+    simpo_gamma: float | None = None
 
 
 class FinetuneRequest(BaseModel):

{together-1.5.13.dist-info → together-1.5.14.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: together
-Version: 1.5.13
+Version: 1.5.14
 Summary: Python client for Together's Cloud Platform!
 License: Apache-2.0
 Author: Together AI

{together-1.5.13.dist-info → together-1.5.14.dist-info}/RECORD

@@ -7,7 +7,7 @@ together/cli/api/chat.py,sha256=2PHRb-9T-lUEKhUJFtc7SxJv3shCVx40gq_8pzfsewM,9234
 together/cli/api/completions.py,sha256=l-Zw5t7hojL3w8xd_mitS2NRB72i5Z0xwkzH0rT5XMc,4263
 together/cli/api/endpoints.py,sha256=f6KafWZvRF6n_ThWdr3y9uhE6wPF37PcD45w_EtgXmY,13289
 together/cli/api/files.py,sha256=QLYEXRkY8J2Gg1SbTCtzGfoTMvosoeACNK83L_oLubs,3397
-together/cli/api/finetune.py,sha256=mM8GF6xP-NM-eOQfn8eXBCLop77OesLVaATAfKm_HMo,16238
+together/cli/api/finetune.py,sha256=zrjxpPSgqcZRhJA4A_QjXNhNUfEu24zw0Da3UfUlzrY,17063
 together/cli/api/images.py,sha256=GADSeaNUHUVMtWovmccGuKc28IJ9E_v4vAEwYHJhu5o,2645
 together/cli/api/models.py,sha256=CXw8B1hqNkadogi58GIXhLg_dTJnvTBaE7Kq1_xQ-10,1423
 together/cli/api/utils.py,sha256=IuqYWPnLI38_Bqd7lj8V_SnGdYc59pRmMbQmciS4FsM,1326
@@ -35,7 +35,7 @@ together/resources/completions.py,sha256=5Wa-ZjPCxRcam6CDe7KgGYlTA7yJZMmd5TrRgGC
 together/resources/embeddings.py,sha256=PTvLb82yjG_-iQOyuhsilp77Fr7gZ0o6WD2KeRnKoxs,2675
 together/resources/endpoints.py,sha256=NNjp-wyzOotzlscGGrANhOHxQBjHTN8f5kTQTH_CLvE,17177
 together/resources/files.py,sha256=y3Ri6UtyAa7fjCJ8_fp26Y2hzzi6Aoo21JKkVgljFl8,5026
-together/resources/finetune.py,sha256=M-nvOZMnL7ZyTayiKN9Vos8D1uYNj-ENmrOA2bkPF8A,37617
+together/resources/finetune.py,sha256=1O8JIbtLDY32N6hL88jUQVDEGcXFnl9qJAEREFoEK5k,40407
 together/resources/images.py,sha256=LQUjKPaFxWTqOAPnyF1Pp7Rz4NLOYhmoKwshpYiprEM,4923
 together/resources/models.py,sha256=qgmAXv61Cq4oLxytenEZBywA8shldDHYxJ_EAu_4JWQ,3864
 together/resources/rerank.py,sha256=3Ju_aRSyZ1s_3zCSNZnSnEJErUVmt2xa3M8z1nvejMA,3931
@@ -52,7 +52,7 @@ together/types/embeddings.py,sha256=J7grkYYn7xhqeKaBO2T-8XQRtHhkzYzymovtGdIUK5A,
 together/types/endpoints.py,sha256=EzNhHOoQ_D9fUdNQtxQPeSWiFzdFLqpNodN0YLmv_h0,4393
 together/types/error.py,sha256=OVlCs3cx_2WhZK4JzHT8SQyRIIqKOP1AZQ4y1PydjAE,370
 together/types/files.py,sha256=i-Ke57p8Svb1MbMZxu-Fo2zxIc6j-mDO2TLGNwPpGu0,1981
-together/types/finetune.py,sha256=Utdcm3kL_cDfBS3zjXwyHsuP2qFFjCQiQZOsPD-WlpE,10918
+together/types/finetune.py,sha256=6_jXgVVp4OOQXkABh0HKBzGy47H3wYCG2QxtXbdYauw,11079
 together/types/images.py,sha256=xnC-FZGdZU30WSFTybfGneWxb-kj0ZGufJsgHtB8j0k,980
 together/types/models.py,sha256=nwQIZGHKZpX9I6mK8z56VW70YC6Ry6JGsVa0s99QVxc,1055
 together/types/rerank.py,sha256=qZfuXOn7MZ6ly8hpJ_MZ7OU_Bi1-cgYNSB20Wja8Qkk,1061
@@ -62,8 +62,8 @@ together/utils/api_helpers.py,sha256=2K0O6qeEQ2zVFvi5NBN5m2kjZJaS3-JfKFecQ7SmGaw
 together/utils/files.py,sha256=btWQawwXbNKfPmCtRyObZViG1Xx-IPz45PrAtMXvcy8,16741
 together/utils/tools.py,sha256=H2MTJhEqtBllaDvOyZehIO_IVNK3P17rSDeILtJIVag,2964
 together/version.py,sha256=p03ivHyE0SyWU4jAnRTBi_sOwywVWoZPU4g2gzRgG-Y,126
-together-1.5.13.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-together-1.5.13.dist-info/METADATA,sha256=pH3dlk_0LIhPs8Z6HW5xnGFPLnxTT7t_-WA3WP4tp1s,15497
-together-1.5.13.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
-together-1.5.13.dist-info/entry_points.txt,sha256=G-b5NKW6lUUf1V1fH8IPTBb7jXnK7lhbX9H1zTEJXPs,50
-together-1.5.13.dist-info/RECORD,,
+together-1.5.14.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+together-1.5.14.dist-info/METADATA,sha256=5fJlYeJKCtS-wVbWPtI_CDWKVSpMDvF3t-DmC8qxZ2U,15497
+together-1.5.14.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
+together-1.5.14.dist-info/entry_points.txt,sha256=G-b5NKW6lUUf1V1fH8IPTBb7jXnK7lhbX9H1zTEJXPs,50
+together-1.5.14.dist-info/RECORD,,