feat(harness): add Action Recognition training experiment

- PerCategoryPoolStrategy: filter test pool by task_types
- RunConfig: add run_holdout_eval toggle (default true)
- load_config: fix YAML task_types list-to-tuple conversion
- Runner: conditionally skip _holdout_four_way when disabled
- CLI: add --no-run-holdout-eval flag
- New config/train_action_recognition.yaml (3 epochs, per_category)
- New scripts/train_action_recognition.sh (baseline + seed + train)

app/harness/config.py

@@ -90,6 +90,7 @@ class RunConfig:
         pool_split_mode: 池划分策略，"global"（全局统一划分）/ "per_category"（按类别独立划分）。
         train_ratio: 训练集占比，范围 (0, 1)。
         test_questions: 测试题目集路径（相对路径）。
+        run_holdout_eval: 是否在 epoch 末执行 held-out 四向评估，默认 True。
     """
 
     # ── 必填字段（无默认值，来自 YAML 或 CLI） ──
@@ -145,6 +146,7 @@ class RunConfig:
     pool_split_mode: str = "global"
     train_ratio: float = 0.667
     test_questions: str = "benchmarks/Video-MME"
+    run_holdout_eval: bool = True
 
 
 def _validate(config: RunConfig) -> None:
@@ -439,6 +441,10 @@ def load_config(
         if field_name in yaml_data:
             yaml_data[field_name] = Path(yaml_data[field_name])
 
+    # Phase 4b: 类型转换 — task_types list → tuple
+    if "task_types" in yaml_data and yaml_data["task_types"] is not None:
+        yaml_data["task_types"] = tuple(yaml_data["task_types"])
+
     # Phase 5: 构造并校验
     config = RunConfig(**{k: v for k, v in yaml_data.items() if k in valid_fields})
     _validate(config)

app/harness/pools.py

@@ -205,6 +205,7 @@ def _q_to_dict(q: GeneratedQuestion) -> dict:
         "answer": q.answer,
         "source_nodes": list(q.source_nodes),
         "difficulty": q.difficulty,
+        "family": q.family,
         "skill_target": q.skill_target,
         "difficulty_steps": q.difficulty_steps,
     }
@@ -228,6 +229,7 @@ def _dict_to_q(d: dict) -> GeneratedQuestion:
         answer=d["answer"],
         source_nodes=tuple(d.get("source_nodes", ())),
         difficulty=d.get("difficulty", "medium"),
+        family=d.get("family"),
         skill_target=d.get("skill_target"),
         difficulty_steps=d.get("difficulty_steps"),
     )
@@ -595,12 +597,15 @@ class PerCategoryPoolStrategy:
             all_train.extend(train)
             all_val.extend(val)
 
-        # Phase 4: test 池（从外部目录加载，无则空）
+        # Phase 4: test 池（从外部目录加载，无则空；按 task_types 过滤）
         test: list[GeneratedQuestion] = []
         if config.test_questions_dir is not None:
             from app.question_gen import load_benchmark
 
             test = load_benchmark(config.test_questions_dir)
+            if config.task_types is not None:
+                allowed = set(config.task_types)
+                test = [q for q in test if q.task_type in allowed]
 
         # Phase 5: 计算 baseline_val_accuracy
         val_correct = sum(1 for q in all_val if correctness.get(q.question_id, False))

app/harness/runner.py

@@ -1368,7 +1368,10 @@ class Runner:
             momentum_updated_task_types=momentum_task_types,
             best_val_acc=state.best_val_acc,
         )
-        await self._holdout_four_way(epoch, pools, state, eval_skills_version, eval_prompts_version)
+        if self._config.run_holdout_eval:
+            await self._holdout_four_way(
+                epoch, pools, state, eval_skills_version, eval_prompts_version
+            )
 
         # Phase 10: gate 阶梯刷新
         self._refresh_gate_ladder(