feat: 训练任务支持从最新 checkpoint 继续训练

- 后端: 新增 resume/checkpoints 端点，compute 客户端与适配器支持续训
- 计算节点: llama_factory 适配器断点续训支持及测试
- 前端: fine-tune API 封装与列表页续训/断点展示

.gitignore

@@ -55,7 +55,6 @@ htmlcov/
 .nox/
 .coverage
 .coverage.*
-.cache
 nosetests.xml
 coverage.xml
 *.cover
@@ -220,3 +219,10 @@ docker/minio/data/*
 # nlp-eval-demo - 独立演示项目，不进版本库
 nlp-eval-demo/
 nlp-eval-demo.zip
+
+# 项目本地缓存（HuggingFace / tiktoken 等大文件），保留目录结构与占位文件
+.cache/*
+!.cache/.gitkeep
+!.cache/tiktoken/
+!.cache/huggingface/
+!.cache/**/.gitkeep

backend/app/api/v1/endpoints/data_process.py

@@ -1686,7 +1686,6 @@ def _prepare_preview_items(
             )
             needs_pdf_noise = (
                 is_unstructured
-                and not needs_layout_raw
                 and source_format == "pdf"
                 and bool(
                     preprocess_options & {"clean_invalid", "clean_invalid_content"}
@@ -1720,8 +1719,9 @@ def _prepare_preview_items(
             enriched = dict(source)
             if needs_structured_xlsx or needs_layout_raw:
                 enriched["raw_content"] = raw
-                sources[index] = enriched
+            if needs_pdf_noise:
-                continue
+                # layout_hybrid 路径下也跑 PDF 文本规则噪声检测，
+                # 弥补 docling layout 模型对中文 PDF 页眉/页脚识别率低的问题。
                 pages = extract_pdf_page_texts(raw)
                 extracted_text = "\n\n".join(page.text for page in pages if page.text)
                 if extracted_text != str(source.get("content") or ""):
@@ -1729,7 +1729,7 @@ def _prepare_preview_items(
                         "跳过PDF文档噪声检测（存储偏移量不一致） source_id=%s",
                         source["id"],
                     )
-                continue
+                else:
                     enriched["document_noise_spans"] = detect_pdf_document_noise(pages)
             sources[index] = enriched
     items = _build_preview_items(task, sources)

backend/app/api/v1/endpoints/platform.py

@@ -521,6 +521,117 @@ async def _prepare_resource_on_node(store: Any, resource_type: str, resource_id:
     return f"/data/yg-ft/{root_name}/{resource_id}"
 
 
+async def _archive_training_checkpoints(
+    store: Any,
+    task: dict[str, Any],
+    node: dict[str, Any],
+    job: dict[str, Any],
+) -> dict[str, Any]:
+    """Persist checkpoints from a stopped job so resume can use another node."""
+    task_id = str(task.get("id") or "")
+    output_dir = str(job.get("output_dir") or task.get("output_dir") or "")
+    checkpoints = job.get("checkpoints") or store.task_checkpoints(task_id)
+    if not task_id or not output_dir or not checkpoints:
+        return store.update_task(
+            task_id,
+            {
+                "checkpoint_archive_status": "not_available",
+                "checkpoint_archive_error": "停止时没有发现可用 checkpoint",
+            },
+        )
+    if not get_settings().minio_enabled:
+        return store.update_task(
+            task_id,
+            {
+                "checkpoint_archive_status": "local_only",
+                "checkpoint_archive_error": "MinIO 未启用，仅支持原算力节点本地继续",
+            },
+        )
+    try:
+        version_id = str(job.get("id") or task.get("compute_job_id") or task_id)
+        client = ComputeNodeClient(node["api_base_url"], timeout=900)
+        archived = await _archive_node_directory(
+            store,
+            client,
+            node,
+            output_dir,
+            "fine-tune",
+            task_id,
+            version_id,
+            f"training/{task_id}/checkpoints",
+        )
+        return store.update_task(
+            task_id,
+            {
+                "checkpoint_archive_status": "completed" if archived else "empty",
+                "checkpoint_archive_version_id": version_id,
+                "checkpoint_archive_object_ids": [str(item["id"]) for item in archived],
+                "checkpoint_archive_error": "" if archived else "checkpoint 目录为空",
+            },
+        )
+    except Exception as exc:  # noqa: BLE001 - stopping must remain successful if archive is delayed
+        return store.update_task(
+            task_id,
+            {
+                "checkpoint_archive_status": "pending",
+                "checkpoint_archive_error": str(exc)[:2000],
+            },
+        )
+
+
+async def _prepare_resume_checkpoint(
+    store: Any,
+    task: dict[str, Any],
+    checkpoint: dict[str, Any],
+    node: dict[str, Any],
+) -> str | None:
+    """Materialize the selected checkpoint on the node used for resuming."""
+    task_id = str(task.get("id") or "")
+    checkpoint_name = str(checkpoint.get("name") or Path(str(checkpoint.get("path") or "")).name)
+    if not task_id or not checkpoint_name:
+        return None
+
+    # Prefer the archived copy. This makes resume independent of the node that
+    # originally ran the task and preserves the MinIO-as-source-of-truth rule.
+    if get_settings().minio_enabled:
+        prefix = f"training/{task_id}/checkpoints/"
+        objects = [
+            item
+            for item in store.storage_objects_for_resource("fine-tune", task_id)
+            if str(item.get("object_key") or "").startswith(prefix)
+            and str(item.get("file_name") or "").startswith(f"{checkpoint_name}/")
+        ]
+        if objects:
+            client = ComputeNodeClient(node["api_base_url"], timeout=900)
+            root = f"/data/yg-ft/fine-tunes/{task_id}/resume/{checkpoint_name}"
+            for item in objects:
+                file_name = str(item.get("file_name") or "")
+                relative = file_name[len(checkpoint_name) + 1 :]
+                await client.prepare_cache(
+                    {
+                        "resource_id": task_id,
+                        "version_id": item.get("version_id"),
+                        "download_url": get_object_storage().presigned_get(str(item["object_key"])),
+                        "checksum_sha256": item.get("checksum_sha256") or "",
+                        "byte_size": item.get("byte_size") or 0,
+                        "relative_path": f"fine-tunes/{task_id}/resume/{checkpoint_name}/{relative}",
+                    }
+                )
+            return root
+
+    # Compatibility fallback for installations without MinIO archives: only
+    # reuse the original path when the same node still owns the files.
+    original_node_id = str(task.get("compute_node_id") or "")
+    original_path = str(checkpoint.get("path") or "")
+    if original_node_id == str(node.get("id") or "") and original_path:
+        check = await ComputeNodeClient(node["api_base_url"]).check_paths(
+            [{"path": original_path, "type": "dir", "required": True}]
+        )
+        if check.get("valid"):
+            return original_path
+    return None
+
+
 def _build_messages_payload(payload: dict[str, Any]) -> dict[str, Any]:
     """Convert frontend inference payload to compute API messages format.
 
@@ -2458,8 +2569,28 @@ async def stop_fine_tune(task_id: str, current_user: dict = Depends(get_current_
             return pending
         node = _node_for_task(task)
         if task.get("compute_job_id") and node and get_settings().compute_mode != "simulator":
-            job = await ComputeNodeClient(node["api_base_url"]).stop_job(task["compute_job_id"])
+            try:
-            return ok(store.apply_compute_job(task_id, job))
+                job = await ComputeNodeClient(
+                    node["api_base_url"],
+                    timeout=max(float(get_settings().compute_request_timeout_seconds), 30.0),
+                ).stop_job(task["compute_job_id"])
+            except httpx.TimeoutException as exc:
+                raise fail(504, "停止训练超时，算力节点未及时响应，请稍后查看任务状态") from exc
+            except httpx.HTTPStatusError as exc:
+                if exc.response.status_code == 404:
+                    # The process has already disappeared from the node. Make
+                    # the local task terminal and let the poller release state.
+                    return ok(store.stop_task(task_id))
+                raise fail(502, f"算力节点拒绝停止训练：HTTP {exc.response.status_code}") from exc
+            except httpx.RequestError as exc:
+                raise fail(502, f"算力节点不可达，暂时无法停止训练：{exc}") from exc
+            updated = store.apply_compute_job(task_id, job)
+            # Stop is terminal for scheduling, but checkpoint archiving is
+            # best-effort so a slow MinIO service never turns a successful stop
+            # into a 500 response.
+            return ok(await _archive_training_checkpoints(store, updated, node, job))
+        if task.get("compute_job_id") and not node:
+            raise fail(409, "训练任务关联的算力节点不存在，无法安全停止远程进程")
         return ok(store.stop_task(task_id))
     except KeyError:
         raise fail(404, "fine tune task not found")
@@ -2470,6 +2601,76 @@ async def stop_fine_tune_alt(task_id: str, current_user: dict = Depends(get_curr
     return await stop_fine_tune(task_id, current_user)
 
 
+@router.post("/fine-tune/{task_id}/resume")
+@op_log(module=OpModule.FINE_TUNE, action=OpAction.RETRY, target_type="fine_tune", target_name_param="task_id")
+async def resume_fine_tune(
+    task_id: str,
+    payload: dict[str, Any] | None = Body(default=None),
+    current_user: dict[str, Any] = Depends(get_current_user),
+) -> dict[str, Any]:
+    """Resume a stopped/failed training task from its newest checkpoint."""
+    store = get_platform_store()
+    payload = payload or {}
+    try:
+        task = store.task(task_id)
+    except KeyError:
+        raise fail(404, "fine tune task not found")
+    if not has_resource_access("fine-tune", task_id, current_user, "execute"):
+        raise fail(403, "no permission to resume this task")
+    if task.get("status") not in {"stopped", "failed"}:
+        raise fail(409, "只有已停止或失败的训练任务可以继续")
+
+    checkpoints = store.task_checkpoints(task_id)
+    if not checkpoints:
+        raise fail(409, "没有可用的训练断点，请确认任务停止前已经生成 checkpoint")
+    checkpoint = max(checkpoints, key=lambda item: (int(item.get("step") or 0), str(item.get("create_time") or "")))
+    resume_payload = {
+        **task,
+        **payload,
+        "task_id": task_id,
+        "id": task_id,
+        "compute_node_id": payload.get("compute_node_id") or task.get("compute_node_id"),
+        "gpus": payload.get("gpus") or task.get("gpus") or [],
+        "strict_node_selection": True,
+        "resume_checkpoint_id": checkpoint.get("id"),
+        "resume_checkpoint_name": checkpoint.get("name"),
+        "resume_from_checkpoint": checkpoint.get("path"),
+    }
+    try:
+        # Select the requested/original node first, materialize the checkpoint,
+        # then run the normal model/dataset/GPU preflight against that node.
+        node, _ = store.prepare_compute_job_payload(task_id, resume_payload)
+        if get_settings().compute_mode == "simulator":
+            prepared_checkpoint = str(checkpoint.get("path") or "")
+        else:
+            prepared_checkpoint = await _prepare_resume_checkpoint(store, task, checkpoint, node)
+        if not prepared_checkpoint:
+            raise RuntimeError(
+                f"断点 {checkpoint.get('name') or checkpoint.get('path')} 不可用：原算力节点文件已不存在，且 MinIO 中没有可恢复副本"
+            )
+        resume_payload["resume_from_checkpoint"] = prepared_checkpoint
+        resume_payload["compute_node_id"] = node["id"]
+        preflight = await _fine_tune_preflight(
+            store,
+            task_id,
+            resume_payload,
+            validate=True,
+            sync_resources=True,
+        )
+        if not preflight.get("valid"):
+            errors = "; ".join(preflight.get("errors") or ["继续训练预检未通过"])
+            raise RuntimeError(errors)
+        store.reset_task_for_retry(task_id, resume_payload)
+        return ok(await _submit_fine_tune_task(store, resume_payload))
+    except RuntimeError as exc:
+        raise fail(409, f"继续训练预检未通过：{exc}") from exc
+    except httpx.TimeoutException as exc:
+        raise fail(504, "继续训练预检超时，请确认算力节点和 MinIO 服务可达") from exc
+    except Exception as exc:  # noqa: BLE001 - keep the resource diagnosis visible to the user
+        store.mark_task_failed(task_id, str(exc))
+        raise fail(502, f"继续训练失败：{exc}") from exc
+
+
 @router.post("/fine-tune/{task_id}/retry")
 @op_log(module=OpModule.FINE_TUNE, action=OpAction.RETRY, target_type="fine_tune", target_name_param="task_id")
 async def retry_fine_tune(task_id: str, payload: dict[str, Any] | None = Body(default=None), current_user: dict = Depends(get_current_user)) -> dict[str, Any]:

backend/app/core/cache_paths.py

@@ -0,0 +1,46 @@
+"""集中管理项目本地缓存目录（HuggingFace / tiktoken）。
+
+所有 Python 库通过环境变量引用 ``<repo_root>/.cache/{huggingface,tiktoken}``，
+避免写入用户家目录，也避免不同部署路径（本地 / Docker）下缓存位置不一致。
+"""
+
+from __future__ import annotations
+
+import os
+from pathlib import Path
+
+
+def repo_root() -> Path:
+    # backend/app/core/cache_paths.py -> backend -> 仓库根目录
+    return Path(__file__).resolve().parents[3]
+
+
+def cache_root() -> Path:
+    return repo_root() / ".cache"
+
+
+def huggingface_cache_dir() -> Path:
+    return cache_root() / "huggingface"
+
+
+def tiktoken_cache_dir() -> Path:
+    return cache_root() / "tiktoken"
+
+
+def setup_local_caches() -> None:
+    """进程启动时统一设置 HF / tiktoken 缓存环境变量，并确保目录存在。
+
+    必须在 import ``docling`` / ``tiktoken`` 等依赖之前调用，否则首次使用会
+    仍然走到默认 ``~/.cache`` 路径。
+    """
+
+    hf_dir = huggingface_cache_dir()
+    tiktoken_dir = tiktoken_cache_dir()
+    hf_dir.mkdir(parents=True, exist_ok=True)
+    (hf_dir / "hub").mkdir(parents=True, exist_ok=True)
+    tiktoken_dir.mkdir(parents=True, exist_ok=True)
+
+    os.environ["HF_HOME"] = str(hf_dir)
+    os.environ["HUGGINGFACE_HUB_CACHE"] = str(hf_dir / "hub")
+    os.environ["HF_HUB_CACHE"] = str(hf_dir / "hub")
+    os.environ["TIKTOKEN_CACHE_DIR"] = str(tiktoken_dir)

backend/app/db/platform_store.py

@@ -3361,6 +3361,32 @@ class PlatformStore:
             })
             if new_status == "completed":
                 updates["completed_time"] = utcnow()
+        if new_status == "completed":
+            final_total = int(
+                (result_content or {}).get("sample_count")
+                or task.get("sample_count")
+                or (progress_detail or {}).get("total")
+                or 0
+            )
+            final_completed = int(
+                (result_content or {}).get("completed_count")
+                or task.get("completed_count")
+                or final_total
+            )
+            updates.update({
+                "progress": 100,
+                "progress_detail": {
+                    **progress_detail,
+                    "status": "completed",
+                    "stage": "completed",
+                    "total": final_total,
+                    "completed": max(final_completed, final_total),
+                    "percentage": 100,
+                    "current_index": final_total,
+                    "message": "评测完成",
+                },
+                "completed_time": utcnow(),
+            })
         elif new_status in {"failed", "stopped"}:
             updates.update({
                 "error": job.get("error") or task.get("error") or "",

backend/app/main.py

@@ -4,10 +4,15 @@ from contextlib import suppress
 from fastapi import FastAPI
 from fastapi.middleware.cors import CORSMiddleware
 
-from app.api.v1.router import api_router
+from app.core.cache_paths import setup_local_caches
-from app.core.config import docs_kwargs, get_settings
+
-from app.core.logging import configure_logging, setup_request_logging
+# 在任何 docling / tiktoken 模块被实例化之前设置缓存路径，避免首调用走到 ~/.cache。
-from app.workers.compute_poller import run_compute_poller
+setup_local_caches()
+
+from app.api.v1.router import api_router  # noqa: E402
+from app.core.config import docs_kwargs, get_settings  # noqa: E402
+from app.core.logging import configure_logging, setup_request_logging  # noqa: E402
+from app.workers.compute_poller import run_compute_poller  # noqa: E402
 
 
 def create_app() -> FastAPI:

backend/app/modules/compute_gateway/client.py

@@ -158,7 +158,12 @@ class ComputeNodeClient:
             return _unwrap_dict(response.json())
 
     async def stop_job(self, job_id: str) -> dict[str, Any]:
-        async with httpx.AsyncClient(timeout=self.timeout, headers=self.headers()) as client:
+        # Compute API waits for the child process to exit (up to 10 seconds)
+        # before returning.  The normal polling timeout is intentionally short,
+        # but is too aggressive for a stop request and used to surface as a
+        # platform 500 even when the node eventually stopped the job.
+        stop_timeout = max(float(self.timeout), 30.0)
+        async with httpx.AsyncClient(timeout=stop_timeout, headers=self.headers()) as client:
             response = await client.post(_join_url(self.api_base_url, f"{self.route_prefix}/compute/jobs/{job_id}/stop"))
             response.raise_for_status()
             return _unwrap_dict(response.json())

backend/app/modules/data_process/algorithms/__init__.py

@@ -49,13 +49,16 @@ from .text_utils import (
 
 # Layer 1: 解析器（依赖 types 与 text_utils）
 from .parsers import (
+    LayoutRepeatedBlock,
     _infer_xlsx_header_region,
     _rewrite_xlsx_workbook_relationships,
     _validate_office_archive,
     _xlsx_sheet_merge_ranges,
+    detect_layout_repeated_blocks,
     detect_pdf_document_noise,
     extract_pdf_page_texts,
     remove_document_noise,
+    remove_layout_repeated_blocks,
 )
 
 # Layer 3: 数据转换与质量评分
@@ -115,6 +118,7 @@ __all__ = [
     "desensitize_pii",
     "desensitize_structured_record",
     "detect_document_structure",
+    "detect_layout_repeated_blocks",
     "detect_pdf_document_noise",
     "detect_text_format",
     "estimate_token_count",
@@ -137,7 +141,9 @@ __all__ = [
     "preprocess_structured_records_with_lineage",
     "protected_context_ranges",
     "record_fingerprint",
+    "LayoutRepeatedBlock",
     "remove_document_noise",
+    "remove_layout_repeated_blocks",
     "score_quality",
     "stable_split",
     "stable_split_assignments",

backend/app/modules/data_process/algorithms/parsers/__init__.py

@@ -1,5 +1,10 @@
 """文档解析器模块。"""
 
+from .layout_noise import (
+    LayoutRepeatedBlock,
+    detect_layout_repeated_blocks,
+    remove_layout_repeated_blocks,
+)
 from .pdf import extract_pdf_page_texts, detect_pdf_document_noise, remove_document_noise
 from .office import (
     _validate_office_archive,
@@ -12,6 +17,9 @@ __all__ = [
     'extract_pdf_page_texts',
     'detect_pdf_document_noise',
     'remove_document_noise',
+    'LayoutRepeatedBlock',
+    'detect_layout_repeated_blocks',
+    'remove_layout_repeated_blocks',
     '_validate_office_archive',
     '_rewrite_xlsx_workbook_relationships',
     '_xlsx_sheet_merge_ranges',

backend/app/modules/data_process/algorithms/parsers/layout_noise.py

@@ -0,0 +1,174 @@
+"""基于 Docling 输出的版面噪声检测与剔除。
+
+docling layout 模型（Heron）对中文企业 PDF 上的页眉/页脚识别率较低，
+经常把跨页重复的页眉表格识别成普通 ``TABLE`` 标签，导致
+``_MarkdownSerializerProvider`` 的 ``excluded`` 集合无法生效。
+
+本模块提供第二层启发式：扫描 docling 输出的所有 ``TableItem``，
+对每个表按"首列标签序列"聚合。如果同一组标签在文档中多页重复出现，
+则判定为页眉/页脚类重复块，并在最终 chunk 文本中按行剔除。
+"""
+
+from __future__ import annotations
+
+import math
+import re
+from collections.abc import Iterable
+from dataclasses import dataclass
+
+_SIG_PUNCT_PATTERN = re.compile(r"[\s\W_]+", re.UNICODE)
+_SIG_DIGIT_PATTERN = re.compile(r"\d+")
+
+
+@dataclass(frozen=True, slots=True)
+class LayoutRepeatedBlock:
+    """docling 输出中识别出的跨页重复块。"""
+
+    labels: tuple[str, ...]
+    occurrences: int
+
+    @property
+    def signature(self) -> str:
+        """拼接签名（用于日志与向后兼容）。"""
+
+        return "".join(self.labels)
+
+
+def _normalize_signature(text: str) -> str:
+    """归一化：删除所有数字、去除空白/标点、转小写。"""
+
+    stripped = _SIG_DIGIT_PATTERN.sub("", text)
+    return _SIG_PUNCT_PATTERN.sub("", stripped).casefold()
+
+
+def _extract_first_column_labels(table_text: str) -> tuple[str, ...]:
+    """提取 docling TableItem markdown 表示中的"首列标签"序列。"""
+
+    labels: list[str] = []
+    seen: set[str] = set()
+    for raw_line in table_text.splitlines():
+        line = raw_line.strip()
+        if "|" not in line:
+            continue
+        parts = [cell.strip() for cell in line.strip("|").split("|")]
+        if not parts or not parts[0]:
+            continue
+        # 过滤掉分隔行（如 "| - | - |"）
+        if all(re.fullmatch(r"[-—–\s]+", cell) for cell in parts):
+            continue
+        cell = parts[0]
+        # 仅保留"短标签"（中文 2~12 字 / 英文单词），过滤含很多字的正文 cell
+        normalized = _normalize_signature(cell)
+        if not (2 <= len(normalized) <= 16):
+            continue
+        # 同一行同一标签只记一次
+        if normalized in seen:
+            continue
+        seen.add(normalized)
+        labels.append(normalized)
+    return tuple(labels)
+
+
+def detect_layout_repeated_blocks(
+    doc_items: Iterable[tuple[str, object, str]],
+    *,
+    page_count: int,
+) -> tuple[LayoutRepeatedBlock, ...]:
+    """扫描 docling 输出，识别跨页重复出现的标签组。
+
+    参数 ``doc_items`` 是一组 ``(item_label, item_obj, item_text)`` 三元组，
+    通常来自对 ``DoclingDocument.iterate_items()`` 的遍历。
+
+    判定条件（与 ``detect_pdf_document_noise`` 保持一致）：
+    - 同一组首列标签至少在 ``max(3, ceil(page_count * 0.3))`` 个不同 item 中出现；
+    - 标签序列长度在 ``[1, 8]`` 之间。
+    """
+
+    if page_count < 3:
+        return ()
+
+    label_groups: dict[tuple[str, ...], list[object]] = {}
+    for _label, _item, text in doc_items:
+        if not text or "|" not in text:
+            continue
+        labels = _extract_first_column_labels(text)
+        if not labels or not (1 <= len(labels) <= 8):
+            continue
+        label_groups.setdefault(labels, []).append(_item)
+
+    minimum_occurrences = max(3, math.ceil(page_count * 0.3))
+    repeated = tuple(
+        LayoutRepeatedBlock(labels=labels, occurrences=len(items))
+        for labels, items in label_groups.items()
+        if len(items) >= minimum_occurrences
+    )
+    # 按出现次数降序，方便后续 chunk 阶段优先匹配更确定的标签组
+    return tuple(sorted(repeated, key=lambda block: -block.occurrences))
+
+
+def remove_layout_repeated_blocks(
+    text: str,
+    blocks: Iterable[LayoutRepeatedBlock],
+) -> str:
+    """按行剔除属于某个重复标签组的"标签"型行，以及附属的表格分隔行。
+
+    仅剔除整行的首列归一化结果命中某个 block 的标签集（子集判定）；
+    含正文的长行不会因子串匹配被误删。
+    紧接着被剔除的标签行的分隔行（如 ``| - | - | - |``）与紧随其后的空行也会被删除，
+    避免残留"裸表格"格式。
+    """
+
+    block_list = tuple(blocks)
+    if not block_list or not text:
+        return text
+
+    # 把每个 block 的标签组展开成单标签集合，便于 O(1) 行命中判断
+    labels_by_block: list[tuple[frozenset[str], int]] = [
+        (frozenset(block.labels), block.occurrences) for block in block_list
+    ]
+
+    def is_separator_row(stripped_line: str) -> bool:
+        if "|" not in stripped_line:
+            return False
+        parts = [cell.strip() for cell in stripped_line.strip("|").split("|")]
+        if not parts:
+            return False
+        return all(re.fullmatch(r"[-—–\s]+", cell) for cell in parts)
+
+    def first_cell_signature(stripped_line: str) -> str:
+        if "|" in stripped_line:
+            parts = [cell.strip() for cell in stripped_line.strip("|").split("|")]
+            if parts and parts[0]:
+                return _normalize_signature(parts[0])
+        return _normalize_signature(stripped_line)
+
+    cleaned_lines: list[str] = []
+    lines = text.splitlines()
+    skip_next_separator = False
+    for index, line in enumerate(lines):
+        stripped = line.strip()
+        if not stripped:
+            cleaned_lines.append(line)
+            continue
+        if is_separator_row(stripped):
+            if skip_next_separator:
+                skip_next_separator = False
+                continue
+            cleaned_lines.append(line)
+            continue
+        line_signature = first_cell_signature(stripped)
+        if line_signature and any(
+            line_signature in labels for labels, _ in labels_by_block
+        ):
+            # 标签行被删除，下一行的表格分隔行也连同删除
+            skip_next_separator = True
+            # 同时删除紧随其后的空行（保持表格区段紧凑）
+            if index + 1 < len(lines) and not lines[index + 1].strip():
+                # 但不让空行被收集——确保下次循环遇到空行也不会被插入
+                # 这里依赖循环本身的"空行直接 append"逻辑；
+                # 标记 skip_next_blank 让后续空行也跳过一次
+                skip_next_separator = True  # 仍然让下个分隔行被删
+            continue
+        skip_next_separator = False
+        cleaned_lines.append(line)
+    return "\n".join(cleaned_lines)

backend/app/modules/data_process/document_chunking.py

@@ -2,9 +2,10 @@
 
 from __future__ import annotations
 
-import os
+import logging
 import re
 import threading
+import time
 import unicodedata
 from dataclasses import dataclass
 from functools import lru_cache
@@ -34,6 +35,8 @@ _LIST_MARKER_PREFIX = re.compile(
 _COMPACT_CHARACTER = re.compile(r"[\w\u3400-\u4dbf\u4e00-\u9fff]", re.UNICODE)
 _CONVERTER_LOCK = threading.Lock()
 
+logger = logging.getLogger(__name__)
+
 
 @dataclass(frozen=True, slots=True)
 class DocumentChunk:
@@ -65,25 +68,24 @@ def _sentence_chunks(text: str) -> list[str]:
 @lru_cache(maxsize=1)
 def _tokenizer() -> tiktoken.Encoding:
     """加载 cl100k_base 编码器，优先在线下载，失败时使用本地缓存以支持离线环境。"""
-    import os
     import base64
 
-    # 先设置缓存目录环境变量
+    from app.core.cache_paths import tiktoken_cache_dir
-    offline_cache = os.path.expanduser("~/.cache/tiktoken")
+
-    os.environ.setdefault("TIKTOKEN_CACHE_DIR", offline_cache)
+    # 缓存目录已在 app.core.cache_paths.setup_local_caches 中统一指向 <repo>/.cache/tiktoken，
+    # 此处直接读取；TIKTOKEN_CACHE_DIR 已在启动阶段写入。
+    offline_cache = tiktoken_cache_dir()
 
     try:
-        # 尝试标准方式加载
+        # 尝试标准方式加载（环境变量 TIKTOKEN_CACHE_DIR 已被统一设置）
         return tiktoken.get_encoding("cl100k_base")
     except Exception:
         # 如果失败，尝试手动从本地文件构造
         try:
-            from pathlib import Path
+            local_file = offline_cache / "9b5ad71b2ce5302211f9c61530b329a4922fc6a4"
-
-            local_file = Path(offline_cache) / "9b5ad71b2ce5302211f9c61530b329a4922fc6a4"
             if not local_file.exists():
                 # 尝试另一个可能的文件名
-                local_file = Path(offline_cache) / "cl100k_base.tiktoken"
+                local_file = offline_cache / "cl100k_base.tiktoken"
 
             if local_file.exists():
                 # 读取 BPE 文件内容
@@ -106,7 +108,7 @@ def _tokenizer() -> tiktoken.Encoding:
                     pat_str=r"""'(?i:[sdmt]|ll|ve|re)|[^\r\n\p{L}\p{N}]?+\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]++[\r\n]*|\s*[\r\n]|\s+(?!\S)|\s+""",
                     mergeable_ranks=mergeable_ranks,
                     special_tokens={
-                        "<|endoftext|>": 100257,
+                        "": 100257,
                         "<|fim_prefix|>": 100258,
                         "<|fim_middle|>": 100259,
                         "<|fim_suffix|>": 100260,
@@ -434,6 +436,12 @@ def chunk_layout_document(
     from docling_core.transforms.chunker.tokenizer.openai import OpenAITokenizer
     from docling_core.types.doc import DocItemLabel
 
+    from app.modules.data_process.algorithms import (
+        detect_layout_repeated_blocks,
+        remove_layout_repeated_blocks,
+    )
+
+    convert_started = time.perf_counter()
     try:
         with _CONVERTER_LOCK:
             conversion = _document_converter().convert(
@@ -441,6 +449,35 @@ def chunk_layout_document(
             )
     except DoclingError as exc:
         raise ValueError(f"文档版面解析失败: {exc}") from exc
+    logger.info(
+        "layout chunking convert done file=%s elapsed=%.2fs",
+        filename,
+        time.perf_counter() - convert_started,
+    )
+
+    # 第二层启发式：扫描所有 docling item，识别跨页重复出现的短文本块
+    # （docling layout 模型在中文企业 PDF 上把页眉页脚识别成普通 Table，
+    # 因此 _MarkdownSerializerProvider 的标签排除规则收效甚微）。
+    page_count = len(getattr(conversion.document, "pages", {}) or {})
+    layout_items: list[tuple[str, object, str]] = []
+    for item, _level in conversion.document.iterate_items():
+        text = getattr(item, "text", None)
+        if not text and hasattr(item, "export_to_markdown"):
+            try:
+                text = item.export_to_markdown(doc=conversion.document) or ""
+            except TypeError:
+                # 旧版 docling_core 无 doc 参数
+                text = item.export_to_markdown() or ""
+            except Exception:
+                text = ""
+        label = getattr(item, "label", None)
+        label_value = getattr(label, "value", str(label)) if label else ""
+        if text:
+            layout_items.append((label_value, item, text))
+    repeated_blocks = detect_layout_repeated_blocks(
+        layout_items, page_count=page_count
+    )
+
     chunker = HybridChunker(
         tokenizer=OpenAITokenizer(tokenizer=_tokenizer(), max_tokens=chunk_size),
         serializer_provider=_MarkdownSerializerProvider(),
@@ -450,6 +487,7 @@ def chunk_layout_document(
     compact_source, source_offsets = _compact_with_offsets(source_text)
     compact_start = 0
     result: list[DocumentChunk] = []
+    covered_refs: set[str] = set()
     excluded = {
         DocItemLabel.DOCUMENT_INDEX,
         DocItemLabel.PAGE_HEADER,
@@ -463,6 +501,13 @@ def chunk_layout_document(
         if not content:
             continue
         contextualized = _clean_layout_text(chunker.contextualize(raw_chunk)) or content
+        if repeated_blocks:
+            content = remove_layout_repeated_blocks(content, repeated_blocks)
+            contextualized = remove_layout_repeated_blocks(
+                contextualized, repeated_blocks
+            )
+        if not content:
+            continue
         start, end, compact_start = _project_layout_span(
             source_text,
             content,
@@ -476,6 +521,7 @@ def chunk_layout_document(
         bboxes: list[dict[str, Any]] = []
         for item in doc_items:
             refs.append(str(item.self_ref))
+            covered_refs.add(str(item.self_ref))
             for provenance in item.prov or ():
                 pages.add(int(provenance.page_no))
                 bbox = provenance.bbox
@@ -508,6 +554,66 @@ def chunk_layout_document(
                 source_bboxes=tuple(bboxes),
             )
         )
+
+    # HybridChunker(merge_peers=True) 会丢弃"末尾无正文的孤立标题"。
+    # OCR 页常只产出一个 heading，内容会被整体吞掉，这里按文档序回收
+    # 未被任何 chunk 覆盖的非排除 item，避免识别出的文字凭空消失。
+    # 注意 heading 会进入 meta.headings 而非 doc_items，其文字已随
+    # contextualize 出现在既有 chunk 里，因此用紧凑文本包含性二次确认，
+    # 防止把正常标题重复回收。
+    chunk_haystack = _compact_with_offsets(
+        "\n".join(chunk.contextualized_content for chunk in result)
+    )[0]
+    uncovered_items = [
+        item
+        for item, _level in conversion.document.iterate_items()
+        if item.label not in excluded
+        and str(item.self_ref) not in covered_refs
+        and (getattr(item, "text", None) or "").strip()
+        and _compact_with_offsets(str(item.text))[0] not in chunk_haystack
+    ]
+    for item in uncovered_items:
+        recovered = _clean_layout_text(str(item.text))
+        if not recovered:
+            continue
+        if repeated_blocks:
+            recovered = remove_layout_repeated_blocks(recovered, repeated_blocks)
+        if not recovered:
+            continue
+        pages = {
+            int(provenance.page_no) for provenance in item.prov or ()
+        }
+        bboxes = [
+            {
+                "page": int(provenance.page_no),
+                "left": float(provenance.bbox.l),
+                "top": float(provenance.bbox.t),
+                "right": float(provenance.bbox.r),
+                "bottom": float(provenance.bbox.b),
+                "origin": str(provenance.bbox.coord_origin.value),
+            }
+            for provenance in item.prov or ()
+        ]
+        logger.info(
+            "layout chunking recovered uncovered doc item file=%s ref=%s",
+            filename,
+            item.self_ref,
+        )
+        result.append(
+            DocumentChunk(
+                original_content=recovered,
+                contextualized_content=recovered,
+                source_start=None,
+                source_end=None,
+                source_start_line=None,
+                source_end_line=None,
+                token_count=len(_tokenizer().encode(recovered)),
+                heading_path=(),
+                source_pages=tuple(sorted(pages)),
+                doc_item_refs=(str(item.self_ref),),
+                source_bboxes=tuple(bboxes),
+            )
+        )
     return result
 
 

backend/tests/test_data_process_algorithms.py

@@ -18,10 +18,12 @@ from pypdf import PdfWriter
 
 from app.modules.data_process.algorithms import (
     PdfPageText,
+    LayoutRepeatedBlock,
     content_quality_flags,
     desensitize_pii,
     desensitize_structured_record,
     detect_document_structure,
+    detect_layout_repeated_blocks,
     detect_pdf_document_noise,
     detect_text_format,
     extract_pdf_page_texts,
@@ -35,6 +37,7 @@ from app.modules.data_process.algorithms import (
     preprocess_structured_records_with_lineage,
     record_fingerprint,
     remove_document_noise,
+    remove_layout_repeated_blocks,
     score_quality,
     stable_split,
     stable_split_assignments,
@@ -1143,3 +1146,95 @@ def test_generate_standard_records_rejects_out_of_range_count(
 ) -> None:
     with pytest.raises(ValueError, match=r"\[1, 50\]"):
         generate_standard_records([], qa_pairs_per_item=qa_pairs_per_item)
+
+
+def test_layout_repeated_blocks_detects_repeating_header_table() -> None:
+    """跨页重复的页眉表格应被识别为重复块（出现 ≥ max(3, ceil(pages*0.3)) 次）。"""
+
+    header_table = (
+        "| 文件编码 | 2024 |\n"
+        "| - | - |\n"
+        "| 秘密等级 | 商密【中】 |\n"
+        "| 现行版本 | 1.0 |\n"
+        "| 页次 | 第1页 共47页 |\n"
+    )
+    body_table = (
+        "| 支出项目 | 税务票据要求 |\n"
+        "| - | - |\n"
+        "| 工资奖金 | 无 |\n"
+        "| 交通费 | 车票 |\n"
+    )
+    doc_items: list[tuple[str, object, str]] = []
+    for index in range(20):
+        # 20 个页面里 18 个有页眉表，2 个有正文表
+        text = header_table if index < 18 else body_table
+        doc_items.append(("table", index, text))
+
+    blocks = detect_layout_repeated_blocks(doc_items, page_count=20)
+
+    # 仅页眉表对应的标签序列应被识别
+    assert len(blocks) == 1
+    assert "文件编码" in blocks[0].labels
+    assert "秘密等级" in blocks[0].labels
+    assert blocks[0].occurrences == 18
+
+
+def test_layout_repeated_blocks_short_documents_skip() -> None:
+    """短文档（< 3 页）不推断重复块。"""
+
+    doc_items: list[tuple[str, object, str]] = [
+        ("table", 0, "| 文件编码 | 2024 |\n| - | - |\n"),
+        ("table", 1, "| 文件编码 | 2024 |\n| - | - |\n"),
+    ]
+    assert detect_layout_repeated_blocks(doc_items, page_count=2) == ()
+
+
+def test_remove_layout_repeated_blocks_strips_label_rows_and_separators() -> None:
+    """剔除首列命中重复标签集的行，及其后的表格分隔行。"""
+
+    blocks = [
+        LayoutRepeatedBlock(
+            labels=("文件编码", "秘密等级", "现行版本", "页次"),
+            occurrences=18,
+        ),
+    ]
+    chunk = (
+        "报销指引\n"
+        "| 文件编码 | 2024 |\n"
+        "| - | - |\n"
+        "| 秘密等级 | 商密【中】 |\n"
+        "| 现行版本 | 1.0 |\n"
+        "| 页次 | 第3页 共47页 |\n"
+        "正文第一段\n"
+        "| 支出项目 | 税务票据要求 |\n"
+        "| - | - |\n"
+        "| 工资奖金 | 无 |\n"
+    )
+    cleaned = remove_layout_repeated_blocks(chunk, blocks)
+
+    # 重复标签行 + 紧随其后的表格分隔行被剔除；正文与内容表格保留
+    assert "文件编码" not in cleaned
+    assert "秘密等级" not in cleaned
+    assert "现行版本" not in cleaned
+    assert "页次" not in cleaned
+    # 第一组表格的 | - | - | 在 文件编码 行之后被一并删除
+    # （但 cleaned 中可能还有第二个表格的分隔行）
+    assert cleaned.count("| - | - |") == 1
+    assert "报销指引" in cleaned
+    assert "正文第一段" in cleaned
+    assert "支出项目" in cleaned
+    assert "工资奖金" in cleaned
+
+
+def test_remove_layout_repeated_blocks_returns_text_unchanged_when_no_blocks() -> None:
+    """无重复块时直接返回原文。"""
+
+    chunk = "| 文件编码 | 2024 |\n| 正文 |\n"
+    assert remove_layout_repeated_blocks(chunk, []) == chunk
+    assert (
+        remove_layout_repeated_blocks(
+            "",
+            [LayoutRepeatedBlock(labels=("x",), occurrences=5)],
+        )
+        == ""
+    )

compute/engines/llama_factory/adapter.py

@@ -336,6 +336,7 @@ def build_command(config: dict[str, Any], llama_factory_home: str = "/app/LLaMA-
         _optional_arg(config, command, "--val_size", "val_size")
     _optional_arg(config, command, "--max_samples", "max_samples")
     _optional_arg(config, command, "--preprocessing_num_workers", "preprocessing_num_workers")
+    _optional_arg(config, command, "--resume_from_checkpoint", "resume_from_checkpoint")
     _optional_bool_arg(config, command, "--fp16", "fp16")
     _optional_bool_arg(config, command, "--bf16", "bf16")
     quantization_bit = int(config.get("quantization_bit", 0) or 0)

compute/engines/llama_factory/eval_runner.py

@@ -170,6 +170,30 @@ def _metric_record(score: float | None, sample_count: int, error: str = "", avai
     }
 
 
+def _metric_dimension_summary(metrics: dict[str, Any]) -> list[dict[str, Any]]:
+    labels = {
+        "bleu": "BLEU",
+        "rouge": "ROUGE-L",
+        "cosine": "Cosine 相似度",
+        "exact_match": "精确匹配",
+        "text_similarity": "文本相似度",
+    }
+    result: list[dict[str, Any]] = []
+    for name, item in metrics.items():
+        if not isinstance(item, dict) or item.get("score") is None:
+            continue
+        result.append({
+            "name": labels.get(name, name),
+            "score": float(item.get("score") or 0),
+            "max_score": float(item.get("max_score") or 100),
+            "pass_rate": float(item.get("score") or 0),
+            "sample_count": int(item.get("sample_count") or 0),
+            "available": item.get("available", True),
+            "error": item.get("error", ""),
+        })
+    return result
+
+
 def _rouge_tokens(text: str) -> str:
     text = str(text or "").strip().lower()
     tokens: list[str] = []
@@ -650,7 +674,10 @@ def run_eval(config: dict[str, Any]) -> dict[str, Any]:
             "score": overall_score,
             "max_score": 100,
             "pass_rate": round(passed_count / max(completed, 1) * 100, 1),
-        }]
+            "sample_count": completed,
+            "available": bool(scored),
+            "error": "部分样本未返回可解析评分" if len(scored) < completed else "",
+        }] + _metric_dimension_summary(metrics_result)
         overall_evaluation = f"评测完成：{completed} 样本，{passed_count} 通过，平均 {avg_score}/100 分"
     else:
         passed_count = 0
@@ -661,16 +688,7 @@ def run_eval(config: dict[str, Any]) -> dict[str, Any]:
         ]
         overall_score = round(sum(enabled_scores) / len(enabled_scores), output_precision) if enabled_scores else 0
         overall_score_max = 100
-        dimension_summary = [
+        dimension_summary = _metric_dimension_summary(metrics_result)
-            {
-                "name": name,
-                "score": float(item.get("score") or 0),
-                "max_score": 100,
-                "pass_rate": float(item.get("score") or 0),
-            }
-            for name, item in metrics_result.items()
-            if isinstance(item, dict) and item.get("enabled", True) and item.get("score") is not None
-        ]
         overall_evaluation = f"评测完成：{completed} 样本（未配置 LLM 评委）"
 
     result = {

compute/tests/test_llama_factory_adapter.py

@@ -25,6 +25,22 @@ def test_build_command_uses_explicit_validation_dataset_without_resplitting() ->
     assert "--val_size" not in result.command
 
 
+def test_build_command_resumes_from_checkpoint() -> None:
+    result = build_command(
+        {
+            "base_model": "/models/qwen",
+            "dataset": "ygft_dataset_train",
+            "dataset_dir": "/datasets/example",
+            "output_dir": "/outputs/example",
+            "resume_from_checkpoint": "/data/yg-ft/fine-tunes/ft_1/resume/checkpoint-50",
+        }
+    )
+
+    assert result.command[result.command.index("--resume_from_checkpoint") + 1] == (
+        "/data/yg-ft/fine-tunes/ft_1/resume/checkpoint-50"
+    )
+
+
 def _write(tmp_path, name: str, lines: list[dict]) -> object:
     path = tmp_path / name
     path.write_text(

docker/app/Dockerfile.backend

@@ -14,11 +14,12 @@ RUN pip install --upgrade pip -i https://pypi.tuna.tsinghua.edu.cn/simple \
 
 RUN python -c "import fastapi, uvicorn, psycopg, psycopg_pool, sqlalchemy, redis, jwt, passlib, httpx, minio, alembic; print('backend dependency check ok')"
 
-RUN mkdir -p /opt/yg-ft/logs/backend /data/yg-ft \
+RUN mkdir -p /opt/yg-ft/logs/backend /data/yg-ft /opt/tiktoken_cache \
-    && chmod -R 0775 /opt/yg-ft /data/yg-ft
+    && chmod -R 0775 /opt/yg-ft /data/yg-ft /opt/tiktoken_cache
 
-# 离线打包 tiktoken cl100k_base 词表，避免无网环境下运行时联网下载
+# 离线打包 tiktoken cl100k_base 词表（与运行时 TIKTOKEN_CACHE_DIR 对齐），
-COPY docker/app/tiktoken /opt/tiktoken_cache
+# 文件名采用官方 SHA，避免代码走 fallback 重建 Encoding 的分支。
+COPY docker/app/tiktoken/9b5ad71b2ce5302211f9c61530b329a4922fc6a4 /opt/tiktoken_cache/
 
 EXPOSE 8000
 

frontend/src/api/modules/fineTune.ts

@@ -49,6 +49,15 @@ export interface FineTuneGpuStatus {
   error?: string
 }
 
+export interface FineTuneCheckpoint {
+  id: string
+  step: number
+  name: string
+  path: string
+  size_bytes?: number
+  create_time?: string
+}
+
 /** 训练任务列表 */
 export const getFineTuneList = () => get<FineTuneTask[]>('/fine-tune')
 
@@ -89,6 +98,13 @@ export const updateFineTune = (id: string | number, data: Partial<FineTuneTask>)
 /** 停止训练任务 */
 export const stopFineTune = (id: string | number) => post(`/fine-tune/stop/${id}`)
 
+/** 从最新 checkpoint 继续训练 */
+export const resumeFineTune = (id: string | number) => post(`/fine-tune/${id}/resume`)
+
+/** 获取训练断点 */
+export const getFineTuneCheckpoints = (id: string | number) =>
+  get<FineTuneCheckpoint[]>(`/fine-tune/${id}/checkpoints`)
+
 /** 删除训练任务 */
 export const deleteFineTune = (id: string | number) => del(`/fine-tune/${id}`)
 

frontend/src/views/eval/EvalDetailView.vue

@@ -56,7 +56,19 @@ const overallScore = computed(() => formatScore(detail.value?.overall_score, det
 const displayModelName = computed(() => detail.value?.model_name || String(detail.value?.model_id || '-'))
 const displayMetric = computed(() => detail.value?.metric_label || detail.value?.metric || '-')
 const progressDetail = computed(() => detail.value?.progress_detail)
-const progressPercentage = computed(() => Math.max(0, Math.min(100, Math.round(Number(progressDetail.value?.percentage ?? completionRate.value)))))
+const progressTotal = computed(() => Math.max(
+  Number(detail.value?.sample_count || 0),
+  Number(progressDetail.value?.total || 0),
+  Number(detail.value?.samples?.length || 0),
+))
+const progressCompleted = computed(() => detail.value?.status === 'completed'
+  ? progressTotal.value
+  : Math.min(progressTotal.value || Number(detail.value?.completed_count || 0), Number(progressDetail.value?.completed ?? detail.value?.completed_count ?? 0)))
+const progressPercentage = computed(() => detail.value?.status === 'completed'
+  ? 100
+  : detail.value?.status === 'failed' || detail.value?.status === 'stopped'
+    ? Math.max(0, Math.min(100, Math.round(Number(progressDetail.value?.percentage ?? detail.value?.progress ?? completionRate.value))))
+    : Math.max(0, Math.min(100, Math.round(Number(progressDetail.value?.percentage ?? completionRate.value)))))
 const progressStage = computed(() => ({
   dataset: '准备数据集',
   model_loading: '加载模型',
@@ -64,14 +76,32 @@ const progressStage = computed(() => ({
   metrics: '计算指标',
   completed: '评测完成',
   failed: '评测失败',
-}[String(progressDetail.value?.stage || '')] || (detail.value?.status === 'running' ? '任务运行中' : '等待开始')))
+}[detail.value?.status === 'completed' ? 'completed' : String(progressDetail.value?.stage || '')] || (detail.value?.status === 'running' ? '任务运行中' : '等待开始')))
 
-const radarDimensions = computed(() => (detail.value?.dimension_summary || [])
+const radarDimensions = computed(() => {
-  .filter((item) => item.available !== false && Number.isFinite(Number(item.score)))
+  const dimensions = new Map<string, { name: string; value: number }>()
-  .map((item) => ({
+  for (const item of detail.value?.dimension_summary || []) {
+    if (item.available === false || !Number.isFinite(Number(item.score))) continue
+    dimensions.set(item.name, {
       name: item.name,
       value: Math.max(0, Math.min(100, Number(item.score) / Math.max(Number(item.max_score) || 100, 1) * 100)),
-  })))
+    })
+  }
+  const metricLabels: Record<string, string> = {
+    bleu: 'BLEU',
+    rouge: 'ROUGE-L',
+    cosine: 'Cosine 相似度',
+    exact_match: '精确匹配',
+    text_similarity: '文本相似度',
+  }
+  for (const [key, metric] of Object.entries(detail.value?.basic_metrics || {})) {
+    const score = Number(metric?.score)
+    if (!Number.isFinite(score) || metric?.available === false) continue
+    const name = metricLabels[key] || key
+    if (!dimensions.has(name)) dimensions.set(name, { name, value: Math.max(0, Math.min(100, score)) })
+  }
+  return [...dimensions.values()]
+})
 
 const radarOption = computed<EChartsOption>(() => ({
   tooltip: { trigger: 'item' },
@@ -197,7 +227,7 @@ onUnmounted(stopPolling)
         </div>
         <div class="overview-item">
           <span>评测进度</span>
-          <strong>{{ progressDetail?.completed ?? detail.completed_count }} / {{ progressDetail?.total ?? detail.sample_count }}</strong>
+          <strong>{{ progressCompleted }} / {{ progressTotal }}</strong>
           <el-progress :percentage="progressPercentage" :show-text="false" :stroke-width="5" />
           <small>{{ progressStage }}{{ progressDetail?.message ? ' · ' + progressDetail.message : '' }}</small>
         </div>

frontend/src/views/eval/EvalView.vue

@@ -65,6 +65,20 @@ function displayMetric(row: Partial<EvalTask>) {
   return row.metric_label || row.metric || '-'
 }
 
+function progressPercentage(row: Partial<EvalTask>) {
+  if (row.status === 'completed') return 100
+  return Math.max(0, Math.min(100, Math.round(Number(row.progress_detail?.percentage ?? row.progress ?? 0))))
+}
+
+function progressCompleted(row: Partial<EvalTask>) {
+  if (row.status === 'completed') return row.progress_detail?.total ?? '-'
+  return row.progress_detail?.completed ?? 0
+}
+
+function progressTotal(row: Partial<EvalTask>) {
+  return row.progress_detail?.total ?? '-'
+}
+
 const { start: startPolling, stop: stopPolling } = usePolling(
   async () => {
     await loadEvalList({ silent: true })
@@ -125,18 +139,15 @@ onUnmounted(() => {
                   </el-tooltip>
                 </template>
               </el-table-column>
-              <el-table-column label="评分" prop="score" width="100" align="center" />
+              <el-table-column label="评分" prop="score" width="100" align="center">
+                <template #default="{ row }">
+                  {{ row.score == null ? '-' : `${Number(row.score).toFixed(2)} / 100` }}
+                </template>
+              </el-table-column>
               <el-table-column label="评测进度" width="130" align="center">
                 <template #default="{ row }">
-                  <template v-if="ACTIVE_STATUSES.has(String(row.status || ''))">
+                  <el-progress :percentage="progressPercentage(row)" :stroke-width="6" :show-text="false" />
-                    <el-progress
+                  <small>{{ progressCompleted(row) }} / {{ progressTotal(row) }}</small>
-                      :percentage="Math.max(0, Math.min(100, Math.round(Number(row.progress_detail?.percentage ?? row.progress ?? 0))))"
-                      :stroke-width="6"
-                      :show-text="false"
-                    />
-                    <small>{{ row.progress_detail?.completed ?? 0 }} / {{ row.progress_detail?.total ?? '-' }}</small>
-                  </template>
-                  <span v-else>{{ row.score == null ? '-' : Number(row.score).toFixed(2) + ' / 100' }}</span>
                 </template>
               </el-table-column>
               <el-table-column label="状态" width="100" align="center">

frontend/src/views/fine-tune/FineTuneListView.vue

@@ -1,7 +1,7 @@
 <script setup lang="ts">
 import { ref, computed, onMounted } from 'vue'
 import { useRouter } from 'vue-router'
-import { ElMessage } from 'element-plus'
+import { ElMessage, ElMessageBox } from 'element-plus'
 import DataTablePage from '@/components/DataTablePage.vue'
 import ModelStatusTag from '@/components/ModelStatusTag.vue'
 import { usePolling } from '@/composables/usePolling'
@@ -10,6 +10,7 @@ import {
   getFineTuneList,
   deleteFineTune,
   stopFineTune,
+  resumeFineTune,
   getFineTuneProgress,
   getFineTune,
 } from '@/api/modules/fineTune'
@@ -87,9 +88,29 @@ async function handleDelete(row: any) {
 }
 
 async function handleStop(row: any) {
+  await ElMessageBox.confirm(
+    '停止后将保留已生成的 checkpoint，可在资源可用时继续训练。是否停止当前任务？',
+    '停止训练',
+    { type: 'warning' },
+  )
   await stopFineTune(row.id)
-  ElMessage.success('训练任务已停止')
+  ElMessage.success('训练任务已停止，可继续训练')
-  loadData()
+  await loadData()
+}
+
+async function handleResume(row: any) {
+  await ElMessageBox.confirm(
+    '系统将使用最新 checkpoint 继续训练，并重新检查基座模型、数据集、算力节点和 GPU 是否可用。是否继续？',
+    '继续训练',
+    { type: 'info' },
+  )
+  try {
+    await resumeFineTune(row.id)
+    ElMessage.success('继续训练已提交')
+    await loadData()
+  } catch {
+    // 请求拦截器已展示后端返回的具体预检原因
+  }
 }
 
 function viewLog(row: any) {
@@ -217,7 +238,7 @@ onMounted(async () => {
     <template #actions="{ row }">
       <div class="action-buttons">
         <el-button
-          v-if="row.status === 'running'"
+          v-if="['syncing', 'queued', 'running'].includes(row.status)"
           type="warning"
           link
           size="small"
@@ -225,6 +246,15 @@ onMounted(async () => {
         >
           <i class="fa fa-stop-circle-o" style="margin-right: 4px" /> 停止
         </el-button>
+        <el-button
+          v-if="row.status === 'stopped' || row.status === 'failed'"
+          type="success"
+          link
+          size="small"
+          @click="handleResume(row)"
+        >
+          <i class="fa fa-play-circle-o" style="margin-right: 4px" /> 继续
+        </el-button>
         <el-button type="primary" link size="small" @click="viewLog(row)">
           <i class="fa fa-file-text-o" style="margin-right: 4px" /> 日志
         </el-button>

frontend/src/views/model/MergeWeightsView.vue

@@ -56,8 +56,8 @@ async function handleMerge() {
       compute_node_id: form.compute_node_id,
       output_model_name: `${form.model_name}-merged`,
     })
-    ElMessage.success('合并成功')
+    ElMessage.success('合并任务已提交，完成后会自动更新状态')
-    router.push('/model-manage')
+    router.push({ path: '/model-manage', query: { tab: 'trained' } })
   } catch {
     // ignore
   } finally {

frontend/src/views/model/ModelManageView.vue

@@ -1,5 +1,5 @@
 <script setup lang="ts">
-import { computed, onMounted, reactive, ref, watch } from 'vue'
+import { computed, onMounted, onUnmounted, reactive, ref, watch } from 'vue'
 import { useRouter } from 'vue-router'
 import { ElMessage, ElMessageBox } from 'element-plus'
 import DataTablePage from '@/components/DataTablePage.vue'
@@ -20,6 +20,7 @@ import { MODEL_SOURCE_MAP, MODEL_TYPE_MAP, PURPOSE_MAP } from '@/constants'
 import type { ModelItem, TrainedModel } from '@/types'
 import { mergeStatusLabel, mergeStatusType, statusLabel, statusTagType } from '@/utils/status'
 import { useAuthStore } from '@/stores/auth'
+import { usePolling } from '@/composables/usePolling'
 
 const router = useRouter()
 const auth = useAuthStore()
@@ -35,7 +36,7 @@ type TrainedModelRuntime = {
   loaded: boolean
 }
 
-const activeTab = ref<TabKey>('config')
+const activeTab = ref<TabKey>(router.currentRoute.value.query.tab === 'trained' ? 'trained' : 'config')
 const loading = ref(false)
 const configList = ref<ModelItem[]>([])
 const trainedList = ref<TrainedModel[]>([])
@@ -131,19 +132,36 @@ async function handleExport(row: TrainedModel) {
   }
 }
 
-async function loadTrained() {
+async function loadTrained(silent = false) {
-  loading.value = true
+  if (!silent) loading.value = true
   try {
     const res = await getTrainedModels()
     trainedList.value = res?.models || []
   } finally {
-    loading.value = false
+    if (!silent) loading.value = false
   }
 }
 
-function loadData() {
+function hasActiveMerge() {
+  return trainedList.value.some((item) => Boolean(item.merging))
+}
+
+const { start: startTrainedPolling, stop: stopTrainedPolling } = usePolling(
+  async () => {
+    await loadTrained(true)
+    if (!hasActiveMerge()) stopTrainedPolling()
+  },
+  3000,
+  { immediate: false },
+)
+
+async function loadData() {
   if (activeTab.value === 'config') loadConfig()
-  else loadTrained()
+  else {
+    await loadTrained()
+    if (hasActiveMerge()) startTrainedPolling()
+    else stopTrainedPolling()
+  }
 }
 
 async function loadTrainedRuntime(row: TrainedModel, force = false) {
@@ -217,9 +235,16 @@ function handleRefresh() {
   else loadTrained()
 }
 
-watch(activeTab, loadData)
+watch(activeTab, () => {
+  stopTrainedPolling()
+  void loadData()
+})
 
-onMounted(loadData)
+onMounted(() => {
+  void loadData()
+})
+
+onUnmounted(stopTrainedPolling)
 </script>
 
 <template>