Add real YOLO train acceptance

2026-06-30 23:33:43 +08:00
parent 5055084788
commit fb96c96d8b
10 changed files with 270 additions and 16 deletions
--- a/README.md
+++ b/README.md
@@ -198,10 +198,11 @@ scripts/check_no_weight_git.sh

 For a fast non-training validation pass, run agents with
 `PYTHONPATH=backend conda run -n seg_smp python scripts/run_agents.py --no-deep`.
-Add `--live`, `--acceptance`, or `--real` only after the backend and frontend
-are running and you want HTTP endpoint, smoke, or real-workspace checks. The
-browser dashboard exposes the same readiness, coverage, GPU, weight, result,
-and agent checks through the UI.
+Add `--live`, `--acceptance`, `--real`, or `--real-train` only after the
+backend and frontend are running and you want HTTP endpoint, smoke,
+real-workspace, or real short-training checks. The browser dashboard exposes
+the same readiness, coverage, GPU, weight, result, and agent checks through
+the UI.

 The web UI includes a dataset bench for creating upload workspaces, uploading
 images/labels/masks, and jumping into the existing rename, PNG conversion,
@@ -272,6 +273,16 @@ real image/mask pair. The latest report is available from
 `GET /api/acceptance/real/latest` and is shown in the coverage panel as
 `真实数据`.

+`POST /api/acceptance/real-train` goes one step further and launches a short
+operator-style YOLO loop on real workspace data. It uploads a real YOLO
+image/txt pair, generates `dataset.yaml`, runs one CPU epoch through
+`yolo.train_custom`, verifies `results.csv` and `best.pt`, then uses that
+trained checkpoint for prediction and GradCAM heatmap jobs. The latest report
+is available from `GET /api/acceptance/real-train/latest` and is shown in the
+coverage panel as `真实短训`. This is heavier than the real-data predict
+acceptance, so run it when you want proof that real uploaded data can create
+loss curves, trained weights, segmentation previews, and heatmap artifacts.
+
 For stronger runtime proof, `POST /api/acceptance/deep` runs minimal training
 loops for the three model families: one SegModel optimizer step, one YOLO
 segmentation epoch on a synthetic 64x64 dataset, one YOLO GradCAM heatmap
@@ -372,8 +383,9 @@ non-training validation pass is needed.

 The web dashboard calls validation in light mode by default:
 `/api/agents/validate?run_build=false&run_acceptance=false&run_deep=false`.
-Pass `run_live=true`, `run_acceptance=true`, `run_real=true`, or
-`run_deep=true` only when you explicitly want the agent to launch live endpoint
-or heavier runtime acceptance checks from the browser/API. Smoke and real data
-acceptance automatically enable the live backend checks because they submit
-jobs through the API.
+Pass `run_live=true`, `run_acceptance=true`, `run_real=true`,
+`run_real_train=true`, or `run_deep=true` only when you explicitly want the
+agent to launch live endpoint or heavier runtime acceptance checks from the
+browser/API. Smoke, real data, and real short-training acceptance
+automatically enable the live backend checks because they submit jobs through
+the API.
--- a/backend/app/acceptance.py
+++ b/backend/app/acceptance.py
@@ -447,6 +447,13 @@ def latest_real_acceptance_report() -> dict[str, Any]:
    return json.loads(path.read_text(encoding="utf-8"))


+def latest_real_train_acceptance_report() -> dict[str, Any]:
+    path = settings.project_root / "var" / "acceptance" / "real_train_latest.json"
+    if not path.exists():
+        return {"available": False, "path": str(path)}
+    return json.loads(path.read_text(encoding="utf-8"))
+
+
 def run_real_dataset_acceptance(base_url: str = "http://127.0.0.1:8010") -> dict[str, Any]:
    """Run the upload/predict/heatmap path against existing non-synthetic Seg data."""
    acceptance_root = settings.project_root / "var" / "acceptance"
@@ -620,6 +627,178 @@ def run_real_dataset_acceptance(base_url: str = "http://127.0.0.1:8010") -> dict
    return report


+def run_real_train_acceptance(base_url: str = "http://127.0.0.1:8010") -> dict[str, Any]:
+    """Run a short YOLO train/predict/heatmap loop using real workspace samples."""
+    acceptance_root = settings.project_root / "var" / "acceptance"
+    run_id = uuid.uuid4().hex[:8]
+    fixture_root = acceptance_root / f"real_train_{run_id}"
+    fixture_root.mkdir(parents=True, exist_ok=True)
+
+    samples = find_real_workspace_samples()
+    checks: list[dict[str, Any]] = [
+        {"name": "real_train_workspace_samples_discovered", "passed": samples["passed"], "detail": samples}
+    ]
+    if not samples["passed"]:
+        report = {
+            "available": True,
+            "run_id": run_id,
+            "base_url": base_url,
+            "fixture_root": str(fixture_root),
+            "passed": False,
+            "checks": checks,
+            "created_at": time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()),
+        }
+        (acceptance_root / "real_train_latest.json").write_text(json.dumps(report, ensure_ascii=False, indent=2), encoding="utf-8")
+        return report
+
+    dataset_name = f"real_train_acceptance_{run_id}"
+    created_dataset = _request_json("POST", f"{base_url}/api/datasets", {"name": dataset_name, "description": "real workspace short train acceptance"}, timeout=10)
+    checks.append({"name": "create_real_train_upload_dataset", "passed": created_dataset.get("passed", False), "detail": created_dataset})
+
+    yolo_image = Path(samples["yolo_pair"]["image"])
+    yolo_label = Path(samples["yolo_pair"]["label"])
+    uploads = {
+        "real_train_yolo_image_upload": _post_file(f"{base_url}/api/datasets/{dataset_name}/upload/images", yolo_image, timeout=30),
+        "real_train_yolo_label_upload": _post_file(f"{base_url}/api/datasets/{dataset_name}/upload/labels", yolo_label, timeout=30),
+    }
+    for name, detail in uploads.items():
+        checks.append({"name": name, "passed": detail.get("passed", False), "detail": detail})
+
+    validation = _request_json("GET", f"{base_url}/api/datasets/{dataset_name}/validate", timeout=20)
+    validation_json = validation.get("json") if validation.get("passed") else {}
+    checks.append(
+        {
+            "name": "real_train_dataset_validate_yolo",
+            "passed": validation.get("passed", False) and validation_json.get("ready", {}).get("yolo"),
+            "detail": validation,
+        }
+    )
+
+    class_count = max(validation_json.get("classes") or [0]) + 1
+    class_names = ["object"] + [f"class_{index}" for index in range(1, class_count)]
+    yolo_yaml = _request_json("POST", f"{base_url}/api/datasets/{dataset_name}/yolo-yaml", {"class_names": class_names}, timeout=20)
+    yolo_yaml_json = yolo_yaml.get("json") if yolo_yaml.get("passed") else {}
+    checks.append({"name": "real_train_dataset_yolo_yaml", "passed": yolo_yaml.get("passed", False), "detail": yolo_yaml})
+
+    train_name = f"{dataset_name}_train"
+    train = _create_job_and_wait(
+        base_url,
+        "yolo.train_custom",
+        {
+            "data": yolo_yaml_json.get("relative_path", f"var/uploads/datasets/{dataset_name}/dataset.yaml"),
+            "model": str(settings.source_root / "Seg_All_In_One_YoloModel" / "yolo11n-seg.pt"),
+            "project": "var/custom_yolo_runs",
+            "name": train_name,
+            "epochs": 1,
+            "imgsz": 96,
+            "batch": 1,
+            "workers": 0,
+            "device": "cpu",
+            "exist_ok": True,
+        },
+        timeout=240,
+    )
+    train_root = settings.project_root / "var" / "custom_yolo_runs" / train_name
+    best_weight = train_root / "weights" / "best.pt"
+    last_weight = train_root / "weights" / "last.pt"
+    results_csv = train_root / "results.csv"
+    checks.append(
+        {
+            "name": "real_train_yolo_one_epoch_job_runner",
+            "passed": train.get("passed", False) and best_weight.exists() and results_csv.exists() and results_csv.stat().st_size > 0,
+            "detail": {
+                **train,
+                "best_weight": _relative_to_project(best_weight) if best_weight.exists() else None,
+                "last_weight": _relative_to_project(last_weight) if last_weight.exists() else None,
+                "results_csv": _relative_to_project(results_csv) if results_csv.exists() else None,
+                "results_csv_size": results_csv.stat().st_size if results_csv.exists() else 0,
+            },
+        }
+    )
+
+    uploaded_image_json = uploads["real_train_yolo_image_upload"].get("json", {})
+    uploaded_image = uploaded_image_json.get("saved", [{}])[0].get("relative_path")
+    predict_name = f"{dataset_name}_predict_trained"
+    if best_weight.exists() and uploaded_image:
+        predict = _create_job_and_wait(
+            base_url,
+            "yolo.predict_custom",
+            {
+                "weights": str(best_weight),
+                "source": uploaded_image,
+                "project": "var/custom_yolo_runs",
+                "name": predict_name,
+                "imgsz": 96,
+                "conf": 0.01,
+                "device": "cpu",
+                "exist_ok": True,
+            },
+            timeout=120,
+        )
+    else:
+        predict = {"passed": False, "error": "skipped because training did not produce best.pt or upload path"}
+    predict_root = settings.project_root / "var" / "custom_yolo_runs" / predict_name
+    predict_outputs = _result_files(predict_root, {".png", ".jpg", ".jpeg"})
+    checks.append(
+        {
+            "name": "real_train_trained_weight_predict_job_runner",
+            "passed": predict.get("passed", False) and bool(predict_outputs),
+            "detail": {**predict, "output_count": len(predict_outputs), "outputs": [_relative_to_project(path) for path in predict_outputs[:8]]},
+        }
+    )
+
+    heatmap_name = f"{dataset_name}_heatmap_trained"
+    if best_weight.exists() and uploaded_image:
+        heatmap = _create_job_and_wait(
+            base_url,
+            "yolo.heatmap_custom",
+            {
+                "weights": str(best_weight),
+                "source": uploaded_image,
+                "project": "var/custom_yolo_runs",
+                "name": heatmap_name,
+                "model_key": "YOLO11n-seg",
+                "pt_name": "best.pt",
+                "cam_method": "GradCAM",
+                "target_layers": "model.model.model[9]",
+                "limit": 1,
+            },
+            timeout=120,
+        )
+    else:
+        heatmap = {"passed": False, "error": "skipped because training did not produce best.pt or upload path"}
+    heatmap_root = settings.project_root / "var" / "custom_yolo_runs" / heatmap_name / "HeartMap_Visual"
+    heatmap_outputs = _result_files(heatmap_root, {".jpg", ".jpeg", ".png"})
+    checks.append(
+        {
+            "name": "real_train_trained_weight_heatmap_job_runner",
+            "passed": heatmap.get("passed", False) and len(heatmap_outputs) >= 2,
+            "detail": {**heatmap, "output_count": len(heatmap_outputs), "outputs": [_relative_to_project(path) for path in heatmap_outputs[:8]]},
+        }
+    )
+
+    report = {
+        "available": True,
+        "run_id": run_id,
+        "base_url": base_url,
+        "fixture_root": str(fixture_root),
+        "dataset_name": dataset_name,
+        "samples": samples,
+        "artifacts": {
+            "train_root": _relative_to_project(train_root),
+            "best_weight": _relative_to_project(best_weight) if best_weight.exists() else None,
+            "results_csv": _relative_to_project(results_csv) if results_csv.exists() else None,
+            "predict_outputs": [_relative_to_project(path) for path in predict_outputs[:8]],
+            "heatmap_outputs": [_relative_to_project(path) for path in heatmap_outputs[:8]],
+        },
+        "passed": all(item["passed"] for item in checks),
+        "checks": checks,
+        "created_at": time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()),
+    }
+    (acceptance_root / "real_train_latest.json").write_text(json.dumps(report, ensure_ascii=False, indent=2), encoding="utf-8")
+    return report
+
+
 def run_deep_acceptance() -> dict[str, Any]:
    """Run minimal training loops for each model family without full datasets."""
    acceptance_root = settings.project_root / "var" / "acceptance"
--- a/backend/app/agents/evaluation_agent.py
+++ b/backend/app/agents/evaluation_agent.py
@@ -108,6 +108,12 @@ def evaluate_project() -> dict:
        and "real_workspace_yolo_predict_job_runner" in acceptance_text
        and "real_workspace_yolo_heatmap_job_runner" in acceptance_text
        and "real_workspace_stack_job_runner" in acceptance_text,
+        "real_train_acceptance": "/api/acceptance/real-train" in backend_text
+        and "runRealTrainAcceptance" in frontend_text
+        and "真实短训" in frontend_text
+        and "real_train_yolo_one_epoch_job_runner" in acceptance_text
+        and "real_train_trained_weight_predict_job_runner" in acceptance_text
+        and "real_train_trained_weight_heatmap_job_runner" in acceptance_text,
        "agent_api": "/api/agents/evaluate" in backend_text and "/api/agents/validate" in backend_text,
        "agent_panel_ui": "runAgentValidation" in frontend_text and "评价建议" in frontend_text and "Validation Agent" in frontend_text,
        "coverage_api": "/api/coverage" in backend_text and coverage["task_build_passed"],
@@ -132,7 +138,7 @@ def evaluate_project() -> dict:
    if coverage["unmapped_user_scripts"]:
        suggestions.append(f"Map remaining user-facing scripts: {len(coverage['unmapped_user_scripts'])}")
    if not suggestions:
-        suggestions.append("Current platform covers the requested control-plane features, uploaded YOLO dataset train/predict/heatmap actions, live uploaded-data YOLO predict/heatmap acceptance, real workspace data acceptance, and synthetic deep training acceptance; next focus is a longer operator-run task on a full dataset.")
+        suggestions.append("Current platform covers the requested control-plane features, uploaded YOLO dataset train/predict/heatmap actions, live uploaded-data YOLO predict/heatmap acceptance, real workspace data acceptance, real short-train acceptance, and synthetic deep training acceptance; next focus is a longer operator-run task on a full dataset.")

    passed_count = sum(1 for item in checks if item["passed"])
    total_count = max(len(checks), 1)
--- a/backend/app/agents/validation_agent.py
+++ b/backend/app/agents/validation_agent.py
@@ -8,7 +8,7 @@ import urllib.error
 import urllib.request
 from pathlib import Path

-from ..acceptance import run_deep_acceptance, run_live_acceptance, run_real_dataset_acceptance
+from ..acceptance import run_deep_acceptance, run_live_acceptance, run_real_dataset_acceptance, run_real_train_acceptance
 from ..capabilities import get_capability_matrix
 from ..catalog import get_catalog
 from ..config import settings
@@ -47,6 +47,7 @@ def validate_project(
    run_acceptance: bool | None = None,
    run_deep: bool | None = None,
    run_real: bool | None = None,
+    run_real_train: bool | None = None,
    run_live: bool | None = None,
 ) -> dict:
    """Validate current runtime readiness without launching heavy training."""
@@ -122,8 +123,9 @@ def validate_project(
    acceptance_enabled = run_acceptance if run_acceptance is not None else os.getenv("SEG_VALIDATE_ACCEPTANCE", "0") == "1"
    deep_enabled = run_deep if run_deep is not None else os.getenv("SEG_VALIDATE_DEEP", "1") == "1"
    real_enabled = run_real if run_real is not None else os.getenv("SEG_VALIDATE_REAL", "0") == "1"
+    real_train_enabled = run_real_train if run_real_train is not None else os.getenv("SEG_VALIDATE_REAL_TRAIN", "0") == "1"
    live_enabled = run_live if run_live is not None else os.getenv("SEG_VALIDATE_LIVE", "0") == "1"
-    live_enabled = live_enabled or acceptance_enabled or real_enabled
+    live_enabled = live_enabled or acceptance_enabled or real_enabled or real_train_enabled

    if live_enabled:
        backend_url = os.getenv("SEG_VALIDATE_BACKEND_URL", "http://127.0.0.1:8010")
@@ -166,6 +168,9 @@ def validate_project(
        if real_enabled:
            real_acceptance = run_real_dataset_acceptance(backend_url)
            checks.append({"name": "real_workspace_acceptance", "passed": real_acceptance["passed"], "detail": real_acceptance})
+        if real_train_enabled:
+            real_train_acceptance = run_real_train_acceptance(backend_url)
+            checks.append({"name": "real_train_acceptance", "passed": real_train_acceptance["passed"], "detail": real_train_acceptance})
        if deep_enabled:
            deep_acceptance = run_deep_acceptance()
            checks.append({"name": "deep_training_acceptance", "passed": deep_acceptance["passed"], "detail": deep_acceptance})
--- a/backend/app/capabilities.py
+++ b/backend/app/capabilities.py
@@ -3,7 +3,12 @@ from __future__ import annotations
 import time
 from typing import Any

-from .acceptance import latest_acceptance_report, latest_deep_acceptance_report, latest_real_acceptance_report
+from .acceptance import (
+    latest_acceptance_report,
+    latest_deep_acceptance_report,
+    latest_real_acceptance_report,
+    latest_real_train_acceptance_report,
+)
 from .catalog import get_catalog
 from .coverage import get_coverage_report
 from .modules.dataset.service import list_uploaded_datasets
@@ -157,6 +162,7 @@ def get_capability_matrix() -> dict[str, Any]:
    gpus = get_gpus()
    acceptance = latest_acceptance_report()
    real_acceptance = latest_real_acceptance_report()
+    real_train_acceptance = latest_real_train_acceptance_report()
    deep_acceptance = latest_deep_acceptance_report()

    all_tasks = catalog["task_types"]
@@ -277,6 +283,12 @@ def get_capability_matrix() -> dict[str, Any]:
            "passed": bool(real_acceptance.get("passed")),
            "detail": real_acceptance.get("run_id", "not run"),
        },
+        {
+            "id": "real_train_acceptance",
+            "label": "真实短训练验收",
+            "passed": bool(real_train_acceptance.get("passed")),
+            "detail": real_train_acceptance.get("run_id", "not run"),
+        },
        {
            "id": "weights_manifest",
            "label": "权重清单",
@@ -302,6 +314,7 @@ def get_capability_matrix() -> dict[str, Any]:
            "gpus_available": bool(gpus.get("available")),
            "acceptance_passed": bool(acceptance.get("passed")),
            "real_acceptance_passed": bool(real_acceptance.get("passed")),
+            "real_train_acceptance_passed": bool(real_train_acceptance.get("passed")),
            "deep_acceptance_passed": bool(deep_acceptance.get("passed")),
        },
        "requirements": requirements,
--- a/backend/app/main.py
+++ b/backend/app/main.py
@@ -13,9 +13,11 @@ from .acceptance import (
    latest_acceptance_report,
    latest_deep_acceptance_report,
    latest_real_acceptance_report,
+    latest_real_train_acceptance_report,
    run_deep_acceptance,
    run_live_acceptance,
    run_real_dataset_acceptance,
+    run_real_train_acceptance,
 )
 from .capabilities import get_capability_matrix
 from .catalog import get_catalog
@@ -131,6 +133,16 @@ def api_real_acceptance(base_url: str = "http://127.0.0.1:8010") -> dict:
    return run_real_dataset_acceptance(base_url)


+@app.get("/api/acceptance/real-train/latest")
+def api_real_train_acceptance_latest() -> dict:
+    return latest_real_train_acceptance_report()
+
+
+@app.post("/api/acceptance/real-train")
+def api_real_train_acceptance(base_url: str = "http://127.0.0.1:8010") -> dict:
+    return run_real_train_acceptance(base_url)
+
+
@app.get("/api/datasets")
 def api_datasets() -> list[dict]:
    return list_uploaded_datasets()
@@ -295,6 +307,7 @@ def api_agent_validate(
    run_acceptance: bool = False,
    run_deep: bool = False,
    run_real: bool = False,
+    run_real_train: bool = False,
    run_live: bool | None = None,
 ) -> dict:
    return validate_project(
@@ -302,5 +315,6 @@ def api_agent_validate(
        run_acceptance=run_acceptance,
        run_deep=run_deep,
        run_real=run_real,
+        run_real_train=run_real_train,
        run_live=run_live,
    )
--- a/backend/tests/test_agents.py
+++ b/backend/tests/test_agents.py
@@ -10,6 +10,7 @@ def test_evaluation_agent_returns_checks():
    assert result["summary"]["passed_checks"] == result["summary"]["total_checks"]
    checks = {item["name"]: item["passed"] for item in result["checks"]}
    assert checks["real_workspace_acceptance"] is True
+    assert checks["real_train_acceptance"] is True


 def test_validation_agent_lightweight(monkeypatch):
--- a/backend/tests/test_capabilities.py
+++ b/backend/tests/test_capabilities.py
@@ -20,3 +20,4 @@ def test_capability_matrix_tracks_user_requirements():
    assert requirements["yolo_heatmap"]["passed"] is True
    assert requirements["training_curves"]["passed"] is True
    assert requirements["real_workspace_acceptance"]["passed"] is True
+    assert requirements["real_train_acceptance"]["passed"] is True
--- a/frontend/src/main.tsx
+++ b/frontend/src/main.tsx
@@ -380,6 +380,7 @@ function useData() {
  const [coverage, setCoverage] = useState<CoveragePayload | null>(null);
  const [acceptance, setAcceptance] = useState<AcceptancePayload | null>(null);
  const [realAcceptance, setRealAcceptance] = useState<AcceptancePayload | null>(null);
+  const [realTrainAcceptance, setRealTrainAcceptance] = useState<AcceptancePayload | null>(null);
  const [deepAcceptance, setDeepAcceptance] = useState<DeepAcceptancePayload | null>(null);
  const [runtimeReadiness, setRuntimeReadiness] = useState<RuntimeReadinessPayload | null>(null);
  const [capabilities, setCapabilities] = useState<CapabilityPayload | null>(null);
@@ -388,7 +389,7 @@ function useData() {

  async function refresh() {
    try {
-      const [catalogNext, gpusNext, envsNext, readinessNext, capabilitiesNext, jobsNext, resultsNext, curvesNext, weightsNext, datasetsNext, coverageNext, acceptanceNext, realAcceptanceNext, deepAcceptanceNext, agentEvaluationNext] = await Promise.all([
+      const [catalogNext, gpusNext, envsNext, readinessNext, capabilitiesNext, jobsNext, resultsNext, curvesNext, weightsNext, datasetsNext, coverageNext, acceptanceNext, realAcceptanceNext, realTrainAcceptanceNext, deepAcceptanceNext, agentEvaluationNext] = await Promise.all([
        api<Catalog>("/api/catalog"),
        api<GpuPayload>("/api/system/gpus"),
        api<CondaEnvPayload>("/api/system/envs"),
@@ -402,6 +403,7 @@ function useData() {
        api<CoveragePayload>("/api/coverage"),
        api<AcceptancePayload>("/api/acceptance/latest"),
        api<AcceptancePayload>("/api/acceptance/real/latest"),
+        api<AcceptancePayload>("/api/acceptance/real-train/latest"),
        api<DeepAcceptancePayload>("/api/acceptance/deep/latest"),
        api<EvaluationAgentPayload>("/api/agents/evaluate")
      ]);
@@ -430,6 +432,7 @@ function useData() {
      setCoverage(coverageNext);
      setAcceptance(acceptanceNext);
      setRealAcceptance(realAcceptanceNext);
+      setRealTrainAcceptance(realTrainAcceptanceNext);
      setDeepAcceptance(deepAcceptanceNext);
      setAgentEvaluation(agentEvaluationNext);
      setError("");
@@ -444,7 +447,7 @@ function useData() {
    return () => window.clearInterval(timer);
  }, []);

-  return { catalog, gpus, condaEnvs, runtimeReadiness, capabilities, agentEvaluation, jobs, results, curves, weightManifest, datasets, datasetValidations, coverage, acceptance, realAcceptance, deepAcceptance, error, refresh };
+  return { catalog, gpus, condaEnvs, runtimeReadiness, capabilities, agentEvaluation, jobs, results, curves, weightManifest, datasets, datasetValidations, coverage, acceptance, realAcceptance, realTrainAcceptance, deepAcceptance, error, refresh };
 }

 function StatusPill({ status }: { status: string }) {
@@ -467,7 +470,7 @@ function JobProgressBar({ progress }: { progress?: JobProgress }) {
 }

 function App() {
-  const { catalog, gpus, condaEnvs, runtimeReadiness, capabilities, agentEvaluation, jobs, results, curves, weightManifest, datasets, datasetValidations, coverage, acceptance, realAcceptance, deepAcceptance, error, refresh } = useData();
+  const { catalog, gpus, condaEnvs, runtimeReadiness, capabilities, agentEvaluation, jobs, results, curves, weightManifest, datasets, datasetValidations, coverage, acceptance, realAcceptance, realTrainAcceptance, deepAcceptance, error, refresh } = useData();
  const [taskType, setTaskType] = useState("mock.echo");
  const [params, setParams] = useState(JSON.stringify(defaultParams["mock.echo"], null, 2));
  const [selectedJob, setSelectedJob] = useState<Job | null>(null);
@@ -671,6 +674,16 @@ function App() {
    }
  }

+  async function runRealTrainAcceptance() {
+    setBusy(true);
+    try {
+      await api("/api/acceptance/real-train", { method: "POST" });
+      await refresh();
+    } finally {
+      setBusy(false);
+    }
+  }
+
  async function runAgentValidation() {
    setAgentBusy(true);
    try {
@@ -1133,6 +1146,9 @@ function App() {
                <button className="iconButton" disabled={busy} onClick={runRealAcceptance} title="运行真实数据验收">
                  <FileSearch size={18} />
                </button>
+                <button className="iconButton" disabled={busy} onClick={runRealTrainAcceptance} title="运行真实短训验收">
+                  <Play size={18} />
+                </button>
                <button className="iconButton" disabled={busy} onClick={runDeepAcceptance} title="运行深度训练验收">
                  <Activity size={18} />
                </button>
@@ -1163,6 +1179,10 @@ function App() {
                <span>真实数据</span>
                <strong>{realAcceptance?.available === false ? "New" : realAcceptance?.passed ? "OK" : "Check"}</strong>
              </div>
+              <div>
+                <span>真实短训</span>
+                <strong>{realTrainAcceptance?.available === false ? "New" : realTrainAcceptance?.passed ? "OK" : "Check"}</strong>
+              </div>
              <div>
                <span>深度训练</span>
                <strong>{deepAcceptance?.available === false ? "New" : deepAcceptance?.passed ? "OK" : "Check"}</strong>
@@ -1174,6 +1194,7 @@ function App() {
                  <span>当前用户侧脚本已全部映射到网页任务。</span>
                  <span>最近验收：{acceptance?.created_at ?? "尚未运行"} {acceptance?.run_id ? `#${acceptance.run_id}` : ""}</span>
                  <span>真实数据：{realAcceptance?.created_at ?? "尚未运行"} {realAcceptance?.run_id ? `#${realAcceptance.run_id}` : ""}，通过 {realAcceptance?.checks?.filter((item) => item.passed).length ?? 0}/{realAcceptance?.checks?.length ?? 0}</span>
+                  <span>真实短训：{realTrainAcceptance?.created_at ?? "尚未运行"} {realTrainAcceptance?.run_id ? `#${realTrainAcceptance.run_id}` : ""}，通过 {realTrainAcceptance?.checks?.filter((item) => item.passed).length ?? 0}/{realTrainAcceptance?.checks?.length ?? 0}</span>
                  <span>深度验收：{deepAcceptance?.created_at ?? "尚未运行"} {deepAcceptance?.run_id ? `#${deepAcceptance.run_id}` : ""}，通过 {deepAcceptance?.checks?.filter((item) => item.passed).length ?? 0}/{deepAcceptance?.checks?.length ?? 0}</span>
                  <span>模型族 readiness：{acceptance?.model_family_readiness?.checks?.filter((item) => item.passed).length ?? 0}/{acceptance?.model_family_readiness?.checks?.length ?? 0}，warnings {acceptance?.model_family_readiness?.warnings?.length ?? 0}</span>
                </>
--- a/scripts/run_agents.py
+++ b/scripts/run_agents.py
@@ -19,6 +19,7 @@ def main() -> None:
    parser.add_argument("--live", action="store_true", help="also check live backend/frontend HTTP endpoints")
    parser.add_argument("--acceptance", action="store_true", help="run the lightweight live acceptance smoke")
    parser.add_argument("--real", action="store_true", help="run real workspace data acceptance through the live backend")
+    parser.add_argument("--real-train", action="store_true", help="run a short real workspace YOLO train/predict/heatmap acceptance")
    parser.add_argument("--no-deep", action="store_true", help="skip synthetic deep training acceptance")
    parser.add_argument("--out", default="var/agent_reports/latest.json")
    args = parser.parse_args()
@@ -29,6 +30,7 @@ def main() -> None:
            run_live=args.live,
            run_acceptance=args.acceptance,
            run_real=args.real,
+            run_real_train=args.real_train,
            run_deep=not args.no_deep,
        ),
    }