fix: route None/degraded diagnoses to lapse; abort on high degrade rate

2026-07-16 06:24:11 -04:00
parent 65126feada
commit efbdeb1647
4 changed files with 91 additions and 4 deletions
@@ -1089,6 +1089,13 @@ class Runner:
            _apply_batch_correctness(state.correctness, log, run_id, batch)

        diagnosis = await self._run_diagnosis(run_id, question_ids=[q.question_id for q in batch])
+        # 降级占比过高疑似 judge 基础设施故障：不以降级信号驱动进化，直接中止
+        n_wrong = sum(1 for q in batch if not state.correctness.get(q.question_id, True))
+        if n_wrong > 0 and diagnosis.degraded_count / n_wrong > 0.5:
+            raise RuntimeError(
+                f"本 step 诊断降级占比 {diagnosis.degraded_count}/{n_wrong} > 50%，"
+                "疑似 judge 基础设施故障，中止训练（不以降级信号驱动进化）。"
+            )
        _accumulate_slow_packs(diagnosis, state)
        await self._gate_batch_skills(epoch, step, diagnosis, total_steps, pools, state)
        # 冷却计数每 step 递减、归零剔除