Skip to content

hle

ipw.evaluation.hle

HLEHandler

Bases: EvaluationHandler

LLM-as-judge evaluation for Humanity's Last Exam (CORRECT / INCORRECT / NOT_ATTEMPTED).

Source code in intelligence-per-watt/src/ipw/evaluation/hle.py 
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
@EvaluationRegistry.register("hle")
class HLEHandler(EvaluationHandler):
    """LLM-as-judge evaluation for Humanity's Last Exam (CORRECT / INCORRECT / NOT_ATTEMPTED)."""

    evaluation_method = "hle"

    def evaluate(
        self,
        *,
        problem: str,
        reference: str,
        model_answer: str,
        metadata: Dict[str, object],
    ) -> Tuple[Optional[bool], Dict[str, object]]:
        if not model_answer or not model_answer.strip():
            return False, {"grade": "NOT_ATTEMPTED", "reason": "empty_response"}

        if not reference or not reference.strip():
            return None, {"reason": "no_ground_truth"}

        if not hasattr(self._client, "chat"):
            raise RuntimeError(
                "HLEHandler requires a client with a .chat() helper (e.g. OpenAIClient)."
            )

        prompt = _GRADER_TEMPLATE.format(
            question=problem,
            target=reference,
            predicted=model_answer,
        )

        try:
            raw = self._client.chat(
                system_prompt="",
                user_prompt=prompt,
                temperature=0.0,
                max_output_tokens=1024,
            )

            grade = _parse_grade(raw)
            is_correct = grade == "CORRECT"

            meta: Dict[str, object] = {
                "grade": grade,
                "raw_judge_output": raw,
            }
            extracted = re.search(
                r"^extracted_final_answer:\s*(.+)", raw, re.MULTILINE
            )
            if extracted:
                meta["extracted_answer"] = extracted.group(1).strip()

            return is_correct, meta

        except Exception as exc:
            LOGGER.error("HLE scoring failed: %s", exc)
            return None, {"error": str(exc)}