Initial commit

verl/workers/reward_model/base.py

@@ -0,0 +1,45 @@
+# Copyright 2024 Bytedance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+The base class for reward model
+"""
+
+from abc import ABC, abstractmethod
+
+from verl import DataProto
+
+
+class BasePPORewardModel(ABC):
+
+    def __init__(self, config):
+        self.config = config
+
+    @abstractmethod
+    def compute_reward(self, data: DataProto) -> DataProto:
+        """Computing reward given input_ids. The transformers should output a tensor with shape
+           [batch_size, sequence_length], and the value at [EOS] mask should be gathered.
+
+        Args:
+            data: must contain keys "input_ids", "attention_mask" and "position_ids".
+                - input_ids: [batch_size, sequence_length]
+                - attention_mask: [batch_size, sequence_length]
+                - position_ids: [batch_size, sequence_length]
+
+        Returns: a data pass protocol containing "reward". Only the [EOS] position contains the reward.
+            Other position should have zero reward. Note that this may change in the future if we use
+            dense reward. So, we leave the interface for general case.
+            - reward: [batch_size, sequence_length].
+
+        """
+        pass