adds AOT (huggingface#1701)

* adds AOT * Applied format changes * added docs and tests --------- Co-authored-by: Igor Melnyk <igor.melnyk@ibm.com>
alexvishnevskiy · Jun 12, 2024 · f5168fd · f5168fd
1 parent 79686e1
commit f5168fd
Show file tree

Hide file tree

Showing 4 changed files with 46 additions and 3 deletions.
diff --git a/docs/source/dpo_trainer.mdx b/docs/source/dpo_trainer.mdx
@@ -121,6 +121,8 @@ The [TR-DPO](https://arxiv.org/pdf/2404.09656) paper suggests syncing the refere
 
 The [RPO](https://arxiv.org/abs/2404.19733) paper implements an iterative preference tuning algorithm using a loss related to the RPO loss in this [paper](https://arxiv.org/abs/2405.16436) that essentially consists of the SFT loss on the chosen preferences together with a weighted DPO loss. To use this loss set the `rpo_alpha` in the `DPOConfig` to an appropriate value.
 
+The [AOT](https://arxiv.org/abs/2406.05882) authors propose to use Distributional Preference Alignment Via Optimal Transport. Traditionally, the alignment algorithms use paired preferences at a sample level, which does not ensure alignment on the distributional level. AOT, on the other hand, can align LLMs on paired or unpaired preference data by making the reward distribution of the positive samples stochastically dominant in the first order on the distribution of negative samples. Specifically, `loss_type="aot"` is appropriate for  paired datasets, where each prompt has both chosen and rejected responses; `loss_type="aot_pair"` is for unpaired datasets. Note that `loss_type="aot_pair"` is similar in spirit to `loss_type="kto_pair"` that applies unpaired alignment methodology on paired dataset. In a nutshell, `loss_type="aot"` ensures that the log-likelihood ratio of chosen to rejected of the aligned model has higher quantiles than that ratio for the reference model. `loss_type="aot_pair"` ensures that the chosen reward is higher on all quantiles than the rejected reward. Note that in both cases quantiles are obtained via sorting. To fully leverage the advantages of the AOT algorithm, it is important to maximize the per-GPU batch size.
+
 ## Logging
 
 While training and evaluating we record the following reward metrics:

diff --git a/tests/test_dpo_trainer.py b/tests/test_dpo_trainer.py
@@ -88,6 +88,10 @@ def _init_dummy_dataset(self):
             ["t5", "ipo", True],
             ["gpt2", "kto_pair", True],
             ["t5", "kto_pair", False],
+            ["gpt2", "aot_pair", True],
+            ["t5", "aot_pair", False],
+            ["gpt2", "aot", True],
+            ["t5", "aot", False],
             ["gpt2", "bco_pair", False],
             ["t5", "bco_pair", True],
             ["gpt2", "sppo_hard", False],
@@ -506,6 +510,14 @@ def test_dpo_lora_bf16_autocast_llama(self):
             ["gpt2", "kto_pair", False, True],
             ["gpt2", "kto_pair", True, False],
             ["gpt2", "kto_pair", True, True],
+            ["gpt2", "aot_pair", False, False],
+            ["gpt2", "aot_pair", False, True],
+            ["gpt2", "aot_pair", True, False],
+            ["gpt2", "aot_pair", True, True],
+            ["gpt2", "aot", False, False],
+            ["gpt2", "aot", False, True],
+            ["gpt2", "aot", True, False],
+            ["gpt2", "aot", True, True],
             ["gpt2", "bco_pair", False, False],
             ["gpt2", "bco_pair", False, True],
             ["gpt2", "bco_pair", True, False],

diff --git a/trl/trainer/dpo_config.py b/trl/trainer/dpo_config.py
@@ -29,7 +29,8 @@ class DPOConfig(TrainingArguments):
             The robust DPO label smoothing parameter from the [cDPO](https://ericmitchell.ai/cdpo.pdf) report and [Robust DPO](https://arxiv.org/abs/2403.00409) paper that should be between 0 and 0.5.
         loss_type (`str`, defaults to `"sigmoid"`):
             The type of DPO loss to use. Either `"sigmoid"` the default DPO loss,`"hinge"` loss from [SLiC](https://arxiv.org/abs/2305.10425) paper, `"ipo"` from [IPO](https://arxiv.org/abs/2310.12036) paper,
-            `"kto_pair"` from the HALOs [report](https://github.com/ContextualAI/HALOs/blob/main/assets/report.pdf), `"bco_pair"` from [BCO](https://arxiv.org/abs/2404.04656) paper or `"robust"` from [Robust DPO](https://arxiv.org/abs/2403.00409) paper.
+            `"kto_pair"` from the HALOs [report](https://github.com/ContextualAI/HALOs/blob/main/assets/report.pdf), `"bco_pair"` from [BCO](https://arxiv.org/abs/2404.04656) paper or `"robust"` from [Robust DPO](https://arxiv.org/abs/2403.00409) paper,
+            "aot" and "aot_pair" from alignment via optimal transport
         label_pad_token_id (`int`, defaults to `-100`):
             The label pad token id. This argument is required if you want to use the default data collator.
         padding_value (`Optional[int]`, *optional*):
@@ -78,7 +79,7 @@ class DPOConfig(TrainingArguments):
     beta: float = 0.1
     label_smoothing: float = 0
     loss_type: Literal[
-        "sigmoid", "hinge", "ipo", "kto_pair", "bco_pair", "sppo_hard", "nca_pair", "robust"
+        "sigmoid", "hinge", "ipo", "kto_pair", "bco_pair", "sppo_hard", "nca_pair", "robust", "aot", "aot_pair"
     ] = "sigmoid"
     label_pad_token_id: int = -100
     padding_value: Optional[int] = None

diff --git a/trl/trainer/dpo_trainer.py b/trl/trainer/dpo_trainer.py
@@ -135,7 +135,7 @@ def __init__(
         ref_model: Optional[Union[PreTrainedModel, nn.Module, str]] = None,
         beta: float = 0.1,
         label_smoothing: float = 0,
-        loss_type: Literal["sigmoid", "hinge", "ipo", "kto_pair", "bco_pair", "robust"] = "sigmoid",
+        loss_type: Literal["sigmoid", "hinge", "ipo", "kto_pair", "bco_pair", "robust", "aot", "aot_pair"] = "sigmoid",
         args: Optional[DPOConfig] = None,
         data_collator: Optional[DataCollator] = None,
         label_pad_token_id: int = -100,
@@ -1066,6 +1066,34 @@ def dpo_loss(
                 - 0.5 * F.logsigmoid(-chosen_rewards)
                 - 0.5 * F.logsigmoid(-rejected_rewards)
             )
+        elif self.loss_type == "aot_pair":
+            chosen_logratios = policy_chosen_logps - reference_chosen_logps
+            rejected_logratios = policy_rejected_logps - reference_rejected_logps
+
+            chosen_logratios_sorted, _ = torch.sort(chosen_logratios, dim=0)
+            rejected_logratios_sorted, _ = torch.sort(rejected_logratios, dim=0)
+
+            delta = chosen_logratios_sorted - rejected_logratios_sorted
+
+            losses = (
+                -F.logsigmoid(self.beta * delta) * (1 - self.label_smoothing)
+                - F.logsigmoid(-self.beta * delta) * self.label_smoothing
+            )
+
+        elif self.loss_type == "aot":
+            pi_logratios = policy_chosen_logps - policy_rejected_logps
+            ref_logratios = reference_chosen_logps - reference_rejected_logps
+
+            pi_logratios_sorted, _ = torch.sort(pi_logratios, dim=0)
+            ref_logratios_sorted, _ = torch.sort(ref_logratios, dim=0)
+
+            delta = pi_logratios_sorted - ref_logratios_sorted
+
+            losses = (
+                -F.logsigmoid(self.beta * delta) * (1 - self.label_smoothing)
+                - F.logsigmoid(-self.beta * delta) * self.label_smoothing
+            )
+
         else:
             raise ValueError(
                 f"Unknown loss type: {self.loss_type}. Should be one of ['sigmoid', 'hinge', 'ipo', 'kto_pair', 'bco_pair', 'sppo_hard', 'nca_pair', 'robust']"