⚖️ Add use_soft_judge option to WinRateCallback (#2347)

* add `use_soft_judge` option to WinRateCallback * formatting * Update trl/trainer/callbacks.py Co-authored-by: Quentin Gallouédec <[email protected]> * renamed soft_win_rate to avg_win_prob * Update trl/trainer/callbacks.py Co-authored-by: Quentin Gallouédec <[email protected]> * fix tests * keep orignal * formatting * Update tests/test_callbacks.py Co-authored-by: Quentin Gallouédec <[email protected]> * Update trl/trainer/callbacks.py Co-authored-by: Quentin Gallouédec <[email protected]> * Update tests/test_callbacks.py Co-authored-by: Quentin Gallouédec <[email protected]> * Update tests/test_callbacks.py * fix test --------- Co-authored-by: Quentin Gallouédec <[email protected]>
huggingface · Nov 15, 2024 · b8c9d9c · b8c9d9c
1 parent 6239631
commit b8c9d9c
Show file tree

Hide file tree

Showing 2 changed files with 75 additions and 6 deletions.
diff --git a/tests/test_callbacks.py b/tests/test_callbacks.py
@@ -30,11 +30,13 @@
 
 
 class HalfPairwiseJudge(BasePairwiseJudge):
-    """Naive pairwise judge that always returns [1, 0]"""
+    """Naive pairwise judge that always returns [1, 0] for two prompts"""
 
-    def judge(self, prompts, completions, shuffle_order=True):
+    def judge(self, prompts, completions, shuffle_order=True, return_scores=False):
         # just check that the batch size is 2
         assert len(prompts) == 2
+        if return_scores:
+            return [0.3, 0.9]
         return [1, 0]
 
 
@@ -132,6 +134,49 @@ def test_without_ref_model(self):
             winrate_history = [h for h in trainer.state.log_history if "eval_win_rate" in h]
             self.assertListEqual(winrate_history, self.expected_winrates)
 
+    def test_soft_judge(self):
+        """Test that the soft judge functionality works correctly"""
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            training_args = TrainingArguments(
+                output_dir=tmp_dir,
+                eval_strategy="steps",
+                eval_steps=2,  # evaluate every 2 steps
+                per_device_train_batch_size=2,  # 8 samples in total so 4 batches of 2 per epoch
+                per_device_eval_batch_size=2,
+                report_to="none",
+            )
+            trainer = TrainerWithRefModel(
+                model=self.model,
+                ref_model=self.ref_model,
+                args=training_args,
+                train_dataset=self.dataset["train"],
+                eval_dataset=self.dataset["test"],
+                processing_class=self.tokenizer,
+            )
+            win_rate_callback = WinRateCallback(
+                judge=self.judge, trainer=trainer, generation_config=self.generation_config, use_soft_judge=True
+            )
+            trainer.add_callback(win_rate_callback)
+            trainer.train()
+
+            # Expected values based on judge returning [0.3, 0.9] for each pair
+            expected_soft_winrates = [
+                {"eval_avg_win_prob": 0.4, "eval_win_rate": 0.5, "epoch": 0.0, "step": 0},
+                {"eval_avg_win_prob": 0.4, "eval_win_rate": 0.5, "epoch": 0.5, "step": 2},
+                {"eval_avg_win_prob": 0.4, "eval_win_rate": 0.5, "epoch": 1.0, "step": 4},
+                {"eval_avg_win_prob": 0.4, "eval_win_rate": 0.5, "epoch": 1.5, "step": 6},
+                {"eval_avg_win_prob": 0.4, "eval_win_rate": 0.5, "epoch": 2.0, "step": 8},
+                {"eval_avg_win_prob": 0.4, "eval_win_rate": 0.5, "epoch": 2.5, "step": 10},
+                {"eval_avg_win_prob": 0.4, "eval_win_rate": 0.5, "epoch": 3.0, "step": 12},
+            ]
+
+            winrate_history = [
+                {k: h[k] for k in ["eval_avg_win_prob", "eval_win_rate", "epoch", "step"]}
+                for h in trainer.state.log_history
+                if "eval_avg_win_prob" in h
+            ]
+            self.assertListEqual(winrate_history, expected_soft_winrates)
+
     @require_peft
     def test_lora(self):
         with tempfile.TemporaryDirectory() as tmp_dir:

diff --git a/trl/trainer/callbacks.py b/trl/trainer/callbacks.py
@@ -230,6 +230,9 @@ class WinRateCallback(TrainerCallback):
             in the evaluation dataset.
         shuffle_order (`bool`, *optional*, defaults to `True`):
             Whether to shuffle the order of the completions before judging.
+        use_soft_judge (`bool`, *optional*, defaults to `False`):
+            Whether to use a soft judge that returns a win probability between 0 and 1 for the first completion vs the
+            second.
     """
 
     def __init__(
@@ -239,12 +242,14 @@ def __init__(
         generation_config: Optional[GenerationConfig] = None,
         num_prompts: Optional[int] = None,
         shuffle_order: bool = True,
+        use_soft_judge: bool = False,
     ):
         self.judge = judge
         self.trainer = trainer
         self.shuffle_order = shuffle_order
         self.generation_config = generation_config
         self.ref_completions = []
+        self.use_soft_judge = use_soft_judge
 
         if self.trainer.eval_dataset is None:
             raise ValueError("Trainer must have an evaluation dataset to use the WinRateCallback.")
@@ -281,15 +286,24 @@ def on_train_begin(self, args: TrainingArguments, state: TrainerState, control:
             )
             # Compute initial win rate as a reference point
             completions = list(zip(self.ref_completions, self.ref_completions))
-            winner_indices = self.judge.judge(prompts, completions, self.shuffle_order)
+            if self.use_soft_judge:
+                ref_win_probs = self.judge.judge(prompts, completions, self.shuffle_order, return_scores=True)
+                winner_indices = [0 if score > 0.5 else 1 for score in ref_win_probs]
+                ref_win_probs = gather_object(ref_win_probs)
+            else:
+                winner_indices = self.judge.judge(prompts, completions, self.shuffle_order)
             prompts = gather_object(prompts)
             completions = gather_object(completions)
             winner_indices = gather_object(winner_indices)
 
         # Logging
         if self.trainer.accelerator.is_main_process:
             win_rate = sum(winner_idx == 1 for winner_idx in winner_indices) / len(winner_indices)
-            self.trainer.log({"eval_win_rate": win_rate})
+            if self.use_soft_judge:
+                avg_win_prob = 1.0 - sum(ref_win_probs) / len(ref_win_probs)
+                self.trainer.log({"eval_avg_win_prob": avg_win_prob, "eval_win_rate": win_rate})
+            else:
+                self.trainer.log({"eval_win_rate": win_rate})
 
             if "wandb" in args.report_to:
                 import wandb
@@ -323,15 +337,25 @@ def on_evaluate(self, args: TrainingArguments, state: TrainerState, control: Tra
             )
 
             completions = list(zip(self.ref_completions, completions))
-            winner_indices = self.judge.judge(prompts, completions, self.shuffle_order)
+
+            if self.use_soft_judge:
+                ref_win_probs = self.judge.judge(prompts, completions, self.shuffle_order, return_scores=True)
+                winner_indices = [0 if score > 0.5 else 1 for score in ref_win_probs]
+                ref_win_probs = gather_object(ref_win_probs)
+            else:
+                winner_indices = self.judge.judge(prompts, completions, self.shuffle_order)
             prompts = gather_object(prompts)
             completions = gather_object(completions)
             winner_indices = gather_object(winner_indices)
 
         # Logging
         if self.trainer.accelerator.is_main_process:
             win_rate = sum(winner_idx == 1 for winner_idx in winner_indices) / len(winner_indices)
-            self.trainer.log({"eval_win_rate": win_rate})
+            if self.use_soft_judge:
+                avg_win_prob = 1.0 - sum(ref_win_probs) / len(ref_win_probs)
+                self.trainer.log({"eval_avg_win_prob": avg_win_prob, "eval_win_rate": win_rate})
+            else:
+                self.trainer.log({"eval_win_rate": win_rate})
 
             if "wandb" in args.report_to:
                 import wandb