sage-eval

SAGE L3 评估组件库，提供可组合的 Metric、Profiler、LLM Judge 实现。

• PyPI: isage-eval
• Import: sage_libs.sage_eval
• Layer: L3 (evaluation library, not benchmark)

边界与约束

• 本仓库只提供评测组件实现，不承载执行平台调度职责。
• 依赖 sage.libs.eval 接口层并在导入时注册实现。
• 采用 fail-fast：不保留 fallback / shim / 兼容分支。

当前组件

Metrics

• AccuracyMetric
• BLEUMetric
• F1Metric

Profilers

• LatencyProfiler
• ThroughputProfiler

Judges

• FaithfulnessJudge
• RelevanceJudge

安装

pip install isage-eval

快速使用

from sage_libs.sage_eval import AccuracyMetric, LatencyProfiler, RelevanceJudge

metric = AccuracyMetric()
metric_result = metric.compute([1, 0, 1], [1, 1, 1])

profiler = LatencyProfiler()
with profiler:
    _ = sum(range(10000))

def mock_llm(_: str) -> str:
    return "SCORE: 0.9\nREASONING: Relevant response."

judge = RelevanceJudge(llm_fn=mock_llm)
judge_result = judge.judge(
    response="Paris is the capital of France.",
    question="What is the capital of France?",
)

License

Apache 2.0