refine

llama_stack/providers/inline/scoring/llm_as_judge/scoring.py

@@ -24,8 +24,9 @@ from llama_stack.providers.utils.common.data_schema_validator import (
 
 from .config import LlmAsJudgeScoringConfig
 from .scoring_fn.llm_as_judge_scoring_fn import LlmAsJudgeScoringFn
+from .scoring_fn.llm_as_judge_math_match_fn import LlmAsJudgeMathMatchFn
 
-LLM_JUDGE_FN = LlmAsJudgeScoringFn
+LLM_JUDGE_FN = [LlmAsJudgeScoringFn, LlmAsJudgeMathMatchFn]
 
 
 class LlmAsJudgeScoringImpl(
@@ -45,8 +46,9 @@ class LlmAsJudgeScoringImpl(
         self.inference_api = inference_api
 
     async def initialize(self) -> None:
-        impl = LLM_JUDGE_FN(inference_api=self.inference_api)
-        self.llm_as_judge_fn = impl
+        for fn in LLM_JUDGE_FN:
+            impl = fn(inference_api=self.inference_api)
+            self.llm_as_judge_fn = impl
 
     async def shutdown(self) -> None: ...
 

llama_stack/providers/inline/scoring/llm_as_judge/scoring_fn/fn_defs/llm_as_judge_405b_math_match.py

@@ -11,7 +11,7 @@ from llama_stack.apis.scoring_functions import (
     ScoringFn,
 )
 
-EQUALITY_TEMPLATE = r"""
+EQUALITY_TEMPLATE = """
 Look at the following two expressions (answers to a math problem) and judge whether they are equivalent. Only perform trivial simplifications
 
 Examples:
@@ -67,8 +67,8 @@ YOUR TASK
 
 Respond with only "Yes" or "No" (without quotes). Do not include a rationale.
 
-    Expression 1: %(expression1)s
-    Expression 2: %(expression2)s
+    Expression 1: {expression1}
+    Expression 2: {expression2}
 """.strip()
 
 
@@ -79,7 +79,7 @@ llm_as_judge_405b_math_match = ScoringFn(
     provider_id="llm-as-judge",
     provider_resource_id="llm-as-judge-405b-math-match",
     params=LLMAsJudgeScoringFnParams(
-        judge_model="meta-llama/Llama-3.1-405B-Instruct",
+        judge_model="openai/gpt-4o",
         prompt_template=EQUALITY_TEMPLATE,
         aggregation_functions=[AggregationFunctionType.accuracy],
     ),

llama_stack/providers/inline/scoring/llm_as_judge/scoring_fn/llm_as_judge_math_match_fn.py

@@ -12,9 +12,9 @@ from llama_stack.providers.utils.scoring.base_scoring_fn import RegisteredBaseSc
 
 from .fn_defs.llm_as_judge_405b_math_match import llm_as_judge_405b_math_match
 from .fn_defs.llm_as_judge_base import llm_as_judge_base
+from ...basic.utils.math_utils import extract_result_from_boxed
 
-
-class LlmAsJudgeScoringFn(RegisteredBaseScoringFn):
+class LlmAsJudgeMathMatchFn(RegisteredBaseScoringFn):
     """
     A scoring_fn that assigns
     """
@@ -47,8 +47,8 @@ class LlmAsJudgeScoringFn(RegisteredBaseScoringFn):
         generated_answer = input_row["generated_answer"]
 
         judge_input_msg = fn_def.params.prompt_template.format(
-            expected_answer=expected_answer,
-            generated_answer=generated_answer,
+            expression1=expected_answer,
+            expression2=extract_result_from_boxed(generated_answer),
         )
 
         print("judge_input_msg", judge_input_msg)
@@ -62,9 +62,11 @@ class LlmAsJudgeScoringFn(RegisteredBaseScoringFn):
             ],
         )
 
-        score = 1.0 if judge_response.lower().strip() == "yes" else 0.0
+        content = judge_response.completion_message.content
+
+        score = 1.0 if content.lower().strip() == "yes" else 0.0
 
         return {
             "score": score,
-            "judge_feedback": judge_response,
+            "judge_feedback": content,
         }

llama_stack/templates/open-benchmark/run.yaml

@@ -219,6 +219,9 @@ benchmarks:
   - benchmark_id: meta-reference-math-500
     dataset_id: math_500
     scoring_functions: ["basic::regex_parser_math_response"]
+  - benchmark_id: meta-reference-math-500-llm-as-judge
+    dataset_id: math_500
+    scoring_functions: ["llm-as-judge::405b-math-match"]
 tool_groups:
 - toolgroup_id: builtin::websearch
   provider_id: tavily-search