chore: 忽略离线部署包，提交安全加固、数据库初始化与文档

- .gitignore: 忽略 docker/offline 离线部署包（镜像/运行时等大文件）
- 安全加固: 新增 compute/api/security.py 及各端安全测试，补充 docs/security-hardening.md
- 数据库: 新增完整初始化 SQL 与 docs/database-config.md
- 数据转换与评测: 修复类型检查、增强校验并补充测试
- Docker 配置与环境变量更新

Co-Authored-By: Claude <noreply@anthropic.com>

backend/app/db/platform_store.py

@@ -239,10 +239,18 @@ def llama_dataset_info(dataset_key: str, file_names: list[str], formatting: str
     fmt = str(formatting).lower()
     for key, file_name in zip(llama_dataset_keys(dataset_key, file_names), file_names):
         if fmt == "sharegpt":
+            # 平台校验按 OpenAI 风格消息（role/content），故 tags 用 role/content
+            # 与 LLaMA-Factory 默认的 from/value 不同，需显式声明避免解析失败。
             result[key] = {
                 "file_name": file_name,
                 "formatting": "sharegpt",
                 "columns": {"messages": "messages"},
+                "tags": {
+                    "role_tag": "role",
+                    "content_tag": "content",
+                    "user_tag": "user",
+                    "assistant_tag": "assistant",
+                },
             }
         elif fmt == "dpo":
             result[key] = {
@@ -273,6 +281,47 @@ def llama_dataset_info(dataset_key: str, file_names: list[str], formatting: str
     return result
 
 
+def _sniff_dataset_format(sample_text: str, max_samples: int = 20) -> str:
+    """嗅探数据集内容格式（兼容 jsonl），返回 sharegpt / dpo / cpt / alpaca。
+
+    按内容而非文件名判断，纯 jsonl 数据集（如 ShareGPT messages、缺省 input 的
+    Alpaca）都能被正确识别，避免训练任务误按 alpaca 解析而失败。
+    """
+    text = (sample_text or "").strip()
+    if not text:
+        return ""
+    records: list[dict[str, Any]] = []
+    try:
+        value = json.loads(text)
+    except (TypeError, ValueError, json.JSONDecodeError):
+        value = None
+    if isinstance(value, list):
+        records = [item for item in value[:max_samples] if isinstance(item, dict)]
+    elif isinstance(value, dict):
+        records = [value]
+    else:
+        for line in text.splitlines()[:max_samples]:
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                obj = json.loads(line)
+            except json.JSONDecodeError:
+                continue
+            if isinstance(obj, dict):
+                records.append(obj)
+    records = records[:max_samples]
+    if not records:
+        return ""
+    if all("messages" in record for record in records):
+        return "sharegpt"
+    if all(record.get("chosen") and record.get("rejected") for record in records):
+        return "dpo"
+    if all(record.get("text") and not (record.get("instruction") or record.get("output")) for record in records):
+        return "cpt"
+    return "alpaca"
+
+
 PASSWORD_HASH_ITERATIONS = 390_000
 
 
@@ -2076,18 +2125,38 @@ class PlatformStore:
         runtime_keys = llama_dataset_keys(dataset_key, runtime_file_names)
         training_keys = runtime_keys[: len(training_files)]
         validation_keys = runtime_keys[len(training_files) :]
-        dataset_format = str(task.get("dataset_format") or (dataset and dataset.get("formatting")) or "alpaca").lower()
-        # P0-2: Validate dataset content against declared format
+        # P0-2: 推导数据集格式并校验内容（兼容 jsonl：按内容嗅探 ShareGPT/DPO/CPT/Alpaca）
         train_type = str(task.get("train_type", task.get("train_method", ""))).upper()
-        expected_format = {
-            "DPO": "dpo",
-            "CPT": "cpt",
-        }.get(train_type)
+        expected_format = {"DPO": "dpo", "CPT": "cpt"}.get(train_type)
+        raw_format = str(
+            task.get("dataset_format")
+            or dataset_metadata.get("format")
+            or (dataset and dataset.get("formatting"))
+            or "alpaca"
+        ).lower()
+        sniffed_format = ""
+        content_samples: dict[str, str] = {}
+        if training_files:
+            with self.connect() as conn:
+                for file_entry in training_files:
+                    sample_row = conn.execute(
+                        "SELECT substr(content, 1, 400000) AS sample FROM dataset_files WHERE id=?",
+                        (str(file_entry["id"]),),
+                    ).fetchone()
+                    sample = (sample_row or {}).get("sample") or ""
+                    content_samples[str(file_entry["id"])] = sample
+                    if not sniffed_format:
+                        sniffed_format = _sniff_dataset_format(sample)
+        known_formats = {"sharegpt", "dpo", "cpt", "pt", "pretrain"}
         if expected_format:
             dataset_format = expected_format
+        elif raw_format in known_formats:
+            dataset_format = raw_format
+        else:
+            dataset_format = sniffed_format or raw_format or "alpaca"
         format_errors: list[str] = []
         for file_entry in training_files:
-            content = file_entry.get("content") or ""
+            content = content_samples.get(str(file_entry["id"])) or ""
             if content:
                 from app.modules.data_process.dataset_format import validate_dataset_format
                 file_errors = validate_dataset_format(dataset_format, content=content)