Making sure only response gets evaluated for log_probs in together.ai…

… api for choice sampling.

PiperOrigin-RevId: 688477264
Change-Id: Ifd14023b4b8128e6cd11d0586b99b34719ac3f72

concordia/language_model/together_ai.py

@@ -49,6 +49,28 @@
 _NUM_INITIAL_TOKENS = 500
 
 
+def _find_response_start_index(tokens):
+  r"""Finds the start of the response in the prompt.
+
+  Args:
+    tokens: A list of strings.
+
+  Returns:
+    The index of the last occurrence of '<start_of_turn>' followed by 'model'
+    and '\n', or 1 if the sequence is not found. This corresponds to the start 
+    of the response.
+  """
+  assert len(tokens) >= 3, "Response doesn't match expectation."
+  for i in range(len(tokens) - 3, -1, -1):
+    if (
+        tokens[i] == '<start_of_turn>'
+        and tokens[i + 1] == 'model'
+        and tokens[i + 2] == '\n'
+    ):
+      return i + 3  # Return the index after the sequence
+  raise ValueError("Response doesn't match expectation.")
+
+
 def _ensure_prompt_not_too_long(
     prompt: str,
     num_response_tokens: int,
@@ -273,8 +295,10 @@ def _sample_choice(
           )
         continue
       else:
-        # removing the first token since it is always scored with None.
-        score = sum(result.prompt[0].logprobs.token_logprobs[1:])
+        logprobs = result.prompt[0].logprobs
+        response_idx = _find_response_start_index(logprobs.tokens)
+        response_log_probs = logprobs.token_logprobs[response_idx:]
+        score = sum(response_log_probs)
         return score
 
     raise language_model.InvalidResponseError(