fix(data-process): 发布精确三路数据切分

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

@@ -534,6 +534,12 @@ async def _sync_training_dataset_to_compute_node(
     files = store.training_dataset_files(dataset_id)
     if not files:
         raise RuntimeError(f"dataset has no uploaded file: {dataset_id}")
+    split_aware = any(item.get("split") for item in files)
+    files = [
+        item
+        for item in files
+        if not split_aware or item.get("split") in {"train", "validation"}
+    ]
     client = ComputeNodeClient(node["api_base_url"])
     results: list[dict[str, Any]] = []
     for item in files:
@@ -1533,4 +1539,3 @@ async def log_content(file: str = Query(...)) -> dict[str, Any]:
 @router.post("/web-log")
 async def web_log(payload: dict[str, Any] = Body(...)) -> dict[str, Any]:
     return ok({"received": True, **payload})
-

backend/app/db/platform_store.py

@@ -1099,11 +1099,28 @@ class PlatformStore:
 
     def _dataset(self, conn: PgConnection, row: PgRow) -> dict[str, Any]:
         files = conn.execute(
-            "SELECT id, name, size, active_version_id, create_time FROM dataset_files WHERE dataset_id=? ORDER BY create_time",
+            """SELECT id, name, size, active_version_id, create_time,
+                      record_count, metadata
+               FROM dataset_files WHERE dataset_id=? ORDER BY create_time, id""",
             (row["id"],),
         ).fetchall()
+        dataset_metadata = json_loads(row.get("metadata"), {})
+        split_counts = dict(dataset_metadata.get("split_counts") or {})
+        if row.get("source") == "task" and not split_counts:
+            split_rows = conn.execute(
+                """SELECT split, COUNT(*) AS count FROM dataset_records
+                   WHERE dataset_id=? GROUP BY split""",
+                (row["id"],),
+            ).fetchall()
+            split_counts = {str(item["split"]): int(item["count"]) for item in split_rows}
         return {
             **dict(row),
+            "metadata": dataset_metadata,
+            "split_counts": {
+                "train": int(split_counts.get("train", 0) or 0),
+                "validation": int(split_counts.get("validation", 0) or 0),
+                "test": int(split_counts.get("test", 0) or 0),
+            },
             "files": [
                 {
                     "id": f["id"],
@@ -1111,6 +1128,8 @@ class PlatformStore:
                     "size": f["size"],
                     "active_version_id": f["active_version_id"],
                     "create_time": f["create_time"],
+                    "record_count": int(f.get("record_count") or 0),
+                    "split": json_loads(f.get("metadata"), {}).get("file_split"),
                 }
                 for f in files
             ],
@@ -1212,14 +1231,22 @@ class PlatformStore:
                 raise KeyError(dataset_id)
             rows = conn.execute(
                 """
-                SELECT id, dataset_id, name, size, content, active_version_id, create_time
+                SELECT id, dataset_id, name, size, content, active_version_id,
+                       create_time, record_count, metadata
                 FROM dataset_files
                 WHERE dataset_id=?
                 ORDER BY create_time
                 """,
                 (dataset_id,),
             ).fetchall()
-            return [dict(row) for row in rows]
+            return [
+                {
+                    **dict(row),
+                    "metadata": json_loads(row.get("metadata"), {}),
+                    "split": json_loads(row.get("metadata"), {}).get("file_split"),
+                }
+                for row in rows
+            ]
 
     def file_versions(self, file_id: str) -> dict[str, Any]:
         row = self.dataset_file(file_id)
@@ -1504,15 +1531,36 @@ class PlatformStore:
             model = conn.execute("SELECT * FROM models WHERE id=?", (base_model_id,)).fetchone()
             dataset = conn.execute("SELECT * FROM datasets WHERE id=?", (dataset_id,)).fetchone()
             files = conn.execute(
-                "SELECT id, name, size, active_version_id, create_time FROM dataset_files WHERE dataset_id=? ORDER BY create_time",
+                """SELECT id, name, size, active_version_id, create_time, metadata
+                   FROM dataset_files WHERE dataset_id=? ORDER BY create_time, id""",
                 (dataset_id,),
             ).fetchall()
         model_path = (model and model.get("path")) or task.get("model_name_or_path") or base_model_id
         if not files:
             raise RuntimeError(f"dataset has no uploaded file: {dataset_id}")
         dataset_key = str(task.get("dataset_key") or llama_dataset_key(dataset_id))
-        dataset_file_names = [Path(str(row["name"] or row["id"])).name for row in files]
+        file_entries = [
-        dataset_keys = llama_dataset_keys(dataset_key, dataset_file_names)
+            {
+                **dict(row),
+                "name": Path(str(row["name"] or row["id"])).name,
+                "split": json_loads(row.get("metadata"), {}).get("file_split"),
+            }
+            for row in files
+        ]
+        split_aware = any(item["split"] for item in file_entries)
+        training_files = [
+            item for item in file_entries if not split_aware or item["split"] == "train"
+        ]
+        validation_files = [
+            item for item in file_entries if split_aware and item["split"] == "validation"
+        ]
+        if not training_files:
+            raise RuntimeError(f"dataset has no training split: {dataset_id}")
+        runtime_files = [*training_files, *validation_files]
+        runtime_file_names = [str(item["name"]) for item in runtime_files]
+        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()
         health_detail = node.get("health_detail") or {}
         dataset_root = str(health_detail.get("dataset_root") or f"{node['data_root'].rstrip('/')}/datasets")
@@ -1526,23 +1574,26 @@ class PlatformStore:
             "name": task["name"],
             "base_model": model_path,
             "model_name_or_path": model_path,
-            "dataset": ",".join(dataset_keys),
+            "dataset": ",".join(training_keys),
             "dataset_key": dataset_key,
-            "dataset_keys": dataset_keys,
+            "dataset_keys": training_keys,
+            "eval_dataset": ",".join(validation_keys) or None,
+            "eval_dataset_keys": validation_keys,
             "dataset_display_name": (dataset and dataset.get("name")) or dataset_id,
             "dataset_dir": dataset_dir,
-            "dataset_info": llama_dataset_info(dataset_key, dataset_file_names, dataset_format),
+            "dataset_info": llama_dataset_info(dataset_key, runtime_file_names, dataset_format),
             "dataset_files": [
                 {
-                    "id": row["id"],
+                    "id": item["id"],
-                    "name": Path(str(row["name"] or row["id"])).name,
+                    "name": item["name"],
-                    "relative_path": f"{dataset_id}/{Path(str(row['name'] or row['id'])).name}",
+                    "relative_path": f"{dataset_id}/{item['name']}",
-                    "local_path": f"{dataset_dir.rstrip('/')}/{Path(str(row['name'] or row['id'])).name}",
+                    "local_path": f"{dataset_dir.rstrip('/')}/{item['name']}",
-                    "active_version_id": row["active_version_id"],
+                    "active_version_id": item["active_version_id"],
-                    "size": row["size"],
+                    "size": item["size"],
-                    "create_time": row["create_time"],
+                    "create_time": item["create_time"],
+                    "split": item["split"],
                 }
-                for row in files
+                for item in runtime_files
             ],
             "output_dir": output_dir,
             "gpus": selected_gpus or task.get("gpus") or [0],
@@ -2536,4 +2587,3 @@ def get_platform_store() -> PlatformStore:
     if _store is None:
         _store = PlatformStore()
     return _store
-

backend/app/modules/data_process/algorithms.py

@@ -2871,6 +2871,51 @@ def stable_split(
     return "test"
 
 
+def stable_split_assignments(
+    values: Sequence[str | int],
+    split: Mapping[str, int] | None = None,
+    *,
+    seed: str = "",
+) -> list[DatasetSplit]:
+    """按稳定顺序和精确配额批量划分数据集。
+
+    单条哈希分桶只能在大样本下近似比例。这里先按哈希稳定排序，再用
+    最大余数法计算各切分配额，确保小数据集也严格遵循配置比例。
+    """
+
+    ratios = dict(split or {"train": 80, "validation": 10, "test": 10})
+    # 复用单条划分的参数校验，避免两套规则逐渐漂移。
+    stable_split("validation", ratios, seed=seed)
+    if not values:
+        return []
+
+    split_order: tuple[DatasetSplit, ...] = ("train", "validation", "test")
+    exact = {name: len(values) * ratios[name] / 100 for name in split_order}
+    quotas = {name: math.floor(exact[name]) for name in split_order}
+    remaining = len(values) - sum(quotas.values())
+    remainder_order = sorted(
+        split_order,
+        key=lambda name: (-(exact[name] - quotas[name]), split_order.index(name)),
+    )
+    for name in remainder_order[:remaining]:
+        quotas[name] += 1
+
+    ranked_indices = sorted(
+        range(len(values)),
+        key=lambda index: (
+            hashlib.sha256(f"{seed}:{values[index]}".encode("utf-8")).digest(),
+            index,
+        ),
+    )
+    assignments: list[DatasetSplit] = ["train"] * len(values)
+    cursor = 0
+    for name in split_order:
+        for index in ranked_indices[cursor : cursor + quotas[name]]:
+            assignments[index] = name
+        cursor += quotas[name]
+    return assignments
+
+
 def _preview_content(item: Mapping[str, Any]) -> str:
     for field in ("edited_content", "editedContent", "original_content", "originalContent", "content"):
         value = item.get(field)
@@ -2970,9 +3015,16 @@ def generate_standard_records(
                     "original_input": input_text,
                     "original_output": output,
                     "status": status,
-                    "split": stable_split(result_id, split, seed=split_seed),
+                    "split": "train",
                 }
             )
+    assignments = stable_split_assignments(
+        [str(result["id"]) for result in results],
+        split,
+        seed=split_seed,
+    )
+    for result, assignment in zip(results, assignments, strict=True):
+        result["split"] = assignment
     return results
 
 
@@ -3021,4 +3073,5 @@ __all__ = [
     "remove_document_noise",
     "score_quality",
     "stable_split",
+    "stable_split_assignments",
 ]

backend/app/modules/data_process/generation.py

@@ -11,7 +11,7 @@ from urllib.parse import urlsplit, urlunsplit
 
 import httpx
 
-from app.modules.data_process.algorithms import normalize_text, stable_split
+from app.modules.data_process.algorithms import normalize_text, stable_split_assignments
 
 
 class ModelGenerationError(ValueError):
@@ -203,7 +203,7 @@ def generate_model_records(
                         "original_output": "",
                         "status": "invalid",
                         "error": error_message,
-                        "split": stable_split(result_id, split, seed=task_id),
+                        "split": "train",
                     }
                 )
                 if on_progress:
@@ -236,7 +236,7 @@ def generate_model_records(
                         "original_output": output,
                         "status": "valid" if valid else "invalid",
                         "error": None if valid else "model result is missing instruction or output",
-                        "split": stable_split(result_id, split, seed=task_id),
+                        "split": "train",
                     }
                 )
             if on_progress:
@@ -244,6 +244,13 @@ def generate_model_records(
     finally:
         if owns_client:
             http_client.close()
+    assignments = stable_split_assignments(
+        [str(result["id"]) for result in results],
+        split,
+        seed=task_id,
+    )
+    for result, assignment in zip(results, assignments, strict=True):
+        result["split"] = assignment
     return results
 
 

backend/app/modules/data_process/store.py

@@ -15,7 +15,7 @@ import psycopg
 from psycopg.rows import dict_row
 
 from app.core.config import get_settings
-from app.modules.data_process.algorithms import estimate_token_count, stable_split
+from app.modules.data_process.algorithms import estimate_token_count, stable_split_assignments
 
 TASK_STATUSES = {"pending", "running", "completed", "failed", "stopped"}
 EDITABLE_STATUSES = {"pending", "failed", "stopped", "completed"}
@@ -1190,15 +1190,15 @@ class DataProcessStore:
         return _decode_row(row) or {}
 
     def publish(self, task_id: str, payload: dict[str, Any]) -> dict[str, Any]:
-        """发布有效结果；任务行锁保证重复请求返回同一数据集。"""
+        """按精确配额发布三个切分文件；重复发布会同步修复既有发布物。"""
         with self.connect() as conn:
             task = self._task_in_connection(conn, task_id, for_update=True)
+            existing_dataset = None
             if task.get("output_dataset_id"):
-                dataset = conn.execute(
+                existing_dataset = conn.execute(
                     "SELECT * FROM datasets WHERE id=%s", (task["output_dataset_id"],)
                 ).fetchone()
-                if dataset:
+                if not existing_dataset:
-                    return {"dataset": _decode_row(dataset), "created": False}
                     # 数据集被外部流程清理后，解除断链并重新发布。
                     conn.execute(
                         "UPDATE data_process_tasks SET output_dataset_id=NULL WHERE id=%s",
@@ -1225,44 +1225,106 @@ class DataProcessStore:
             if invalid_count:
                 raise InvalidStateError(f"task contains {invalid_count} invalid results")
 
-            dataset_id = new_id("dataset")
+            dataset_id = (
-            file_id = new_id("dfile")
+                str(existing_dataset["id"]) if existing_dataset else new_id("dataset")
-            version_id = new_id("dfv")
+            )
             now = utcnow()
             requested_split = payload.get("split") or {
                 "train": 80,
                 "validation": 10,
                 "test": 10,
             }
+            assignments = stable_split_assignments(
+                [str(row["id"]) for row in rows],
+                requested_split,
+                seed=task_id,
+            )
             records = [
                 {
                     "instruction": row["instruction"],
                     "input": row["input"],
                     "output": row["output"],
-                    "split": stable_split(
+                    "split": assignment,
-                        str(row["id"]),
+                }
-                        requested_split,
+                for row, assignment in zip(rows, assignments, strict=True)
-                        seed=task_id,
+            ]
+            split_order = ("train", "validation", "test")
+            split_counts = {
+                split_name: assignments.count(split_name) for split_name in split_order
+            }
+            file_specs: list[dict[str, Any]] = []
+            for split_name in split_order:
+                split_records = [
+                    (source_row, record)
+                    for source_row, record in zip(rows, records, strict=True)
+                    if record["split"] == split_name
+                ]
+                if not split_records:
+                    continue
+                file_id = new_id("dfile")
+                version_id = new_id("dfv")
+                content = "".join(
+                    json_dumps(record) + "\n" for _, record in split_records
+                )
+                raw = content.encode("utf-8")
+                file_specs.append(
+                    {
+                        "split": split_name,
+                        "records": split_records,
+                        "file_id": file_id,
+                        "version_id": version_id,
+                        "content": content,
+                        "raw": raw,
+                        "checksum": hashlib.sha256(raw).hexdigest(),
+                        "storage_object_id": (
+                            f"db://data-process/{task_id}/{file_id}/v1"
                         ),
                     }
-                for row in rows
+                )
-            ]
+            total_size = sum(len(spec["raw"]) for spec in file_specs)
-            content = "".join(json_dumps(record) + "\n" for record in records)
-            raw = content.encode("utf-8")
-            checksum = hashlib.sha256(raw).hexdigest()
-            storage_object_id = f"db://data-process/{task_id}/{file_id}/v1"
             source_result_ids = [row["id"] for row in rows]
             metadata = {
                 "source": "data_process",
                 "storage_backend": "database",
-                "storage_object_id": storage_object_id,
                 "source_task_id": task_id,
                 "source_file_ids": [item["id"] for item in self._source_ids(conn, task_id)],
                 "source_result_ids": source_result_ids,
                 "format": payload.get("format") or "alpaca_jsonl",
-                "split": payload.get("split") or {},
+                "split": requested_split,
+                "split_counts": split_counts,
             }
             try:
+                if existing_dataset:
+                    # 用户显式重新发布时，在同一数据集 ID 下修复旧的混合文件。
+                    conn.execute(
+                        "DELETE FROM dataset_records WHERE dataset_id=%s", (dataset_id,)
+                    )
+                    conn.execute(
+                        """DELETE FROM dataset_file_versions
+                           WHERE dataset_file_id IN
+                           (SELECT id FROM dataset_files WHERE dataset_id=%s)""",
+                        (dataset_id,),
+                    )
+                    conn.execute("DELETE FROM dataset_files WHERE dataset_id=%s", (dataset_id,))
+                    dataset = conn.execute(
+                        """
+                        UPDATE datasets
+                        SET size=%s, size_bytes=%s, count=%s, record_count=%s,
+                            description=%s, metadata=%s, updated_at=%s
+                        WHERE id=%s RETURNING *
+                        """,
+                        (
+                            f"{total_size} B",
+                            total_size,
+                            len(records),
+                            len(records),
+                            payload.get("description") or task.get("description") or "",
+                            json_dumps(metadata),
+                            now,
+                            dataset_id,
+                        ),
+                    ).fetchone()
+                else:
                     dataset = conn.execute(
                         """
                         INSERT INTO datasets
@@ -1281,8 +1343,8 @@ class DataProcessStore:
                             payload.get("storage_type") or "local",
                             task_id,
                             task_id,
-                        f"{len(raw)} B",
+                            f"{total_size} B",
-                        len(raw),
+                            total_size,
                             len(records),
                             len(records),
                             payload.get("description") or task.get("description") or "",
@@ -1296,16 +1358,21 @@ class DataProcessStore:
                             now,
                         ),
                     ).fetchone()
+
+                for spec in file_specs:
+                    file_metadata = {**metadata, "file_split": spec["split"]}
                     version = {
-                    "id": version_id,
+                        "id": spec["version_id"],
                         "version_no": 1,
-                    "description": "data process publish",
+                        "version": 1,
-                    "checksum_sha256": checksum,
+                        "description": f"data process {spec['split']} publish",
-                    "size_bytes": len(raw),
+                        "checksum_sha256": spec["checksum"],
-                    "record_count": len(records),
+                        "size_bytes": len(spec["raw"]),
+                        "record_count": len(spec["records"]),
                         "created_at": now,
+                        "create_time": now,
                         "source_task_id": task_id,
-                    "storage_object_id": storage_object_id,
+                        "storage_object_id": spec["storage_object_id"],
                     }
                     conn.execute(
                         """
@@ -1319,25 +1386,25 @@ class DataProcessStore:
                                 %s, %s, 1, %s, %s, %s, %s, %s, %s, %s)
                         """,
                         (
-                        file_id,
+                            spec["file_id"],
                             dataset_id,
-                        f"{payload['dataset_name']}.jsonl",
+                            f"{payload['dataset_name']}.{spec['split']}.jsonl",
-                        storage_object_id,
+                            spec["storage_object_id"],
-                        f"{len(raw)} B",
+                            f"{len(spec['raw'])} B",
-                        content,
+                            spec["content"],
-                        version_id,
+                            spec["version_id"],
                             json_dumps([version]),
                             now,
-                        version_id,
+                            spec["version_id"],
-                        len(raw),
+                            len(spec["raw"]),
-                        len(records),
+                            len(spec["records"]),
                             "jsonl",
-                        checksum,
+                            spec["checksum"],
                             task_id,
                             task.get("tenant_id"),
                             task.get("project_id"),
                             payload.get("created_by") or task.get("created_by"),
-                        json_dumps(metadata),
+                            json_dumps(file_metadata),
                             now,
                             now,
                         ),
@@ -1351,22 +1418,22 @@ class DataProcessStore:
                         VALUES (%s, %s, 1, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s)
                         """,
                         (
-                        version_id,
+                            spec["version_id"],
-                        file_id,
+                            spec["file_id"],
-                        storage_object_id,
+                            spec["storage_object_id"],
-                        content[:2000],
+                            spec["content"][:2000],
-                        "data process publish",
+                            f"data process {spec['split']} publish",
-                        len(raw),
+                            len(spec["raw"]),
-                        len(records),
+                            len(spec["records"]),
-                        checksum,
+                            spec["checksum"],
                             task_id,
-                        json_dumps(metadata),
+                            json_dumps(file_metadata),
                             payload.get("created_by") or task.get("created_by"),
                             now,
                         ),
                     )
                     for line_number, (source_row, record) in enumerate(
-                    zip(rows, records, strict=True), start=1
+                        spec["records"], start=1
                     ):
                         conn.execute(
                             """
@@ -1380,8 +1447,8 @@ class DataProcessStore:
                             (
                                 new_id("drec"),
                                 dataset_id,
-                            file_id,
+                                spec["file_id"],
-                            version_id,
+                                spec["version_id"],
                                 line_number,
                                 record["split"],
                                 record["instruction"],
@@ -1412,7 +1479,11 @@ class DataProcessStore:
                 """,
                 (dataset_id, now, payload.get("created_by"), task_id),
             )
-        return {"dataset": _decode_row(dataset), "created": True}
+        return {
+            "dataset": _decode_row(dataset),
+            "created": existing_dataset is None,
+            "split_counts": split_counts,
+        }
 
     @staticmethod
     def _source_ids(

backend/tests/test_data_process_algorithms.py

@@ -35,6 +35,7 @@ from app.modules.data_process.algorithms import (
     remove_document_noise,
     score_quality,
     stable_split,
+    stable_split_assignments,
 )
 
 
@@ -859,6 +860,17 @@ def test_stable_split_is_reproducible_and_validates_ratios() -> None:
         stable_split("record", {"train": 80, "validation": 10, "test": 9})
 
 
+def test_stable_split_assignments_use_exact_deterministic_quotas() -> None:
+    values = [f"record-{index}" for index in range(28)]
+    first = stable_split_assignments(values, seed="task-1")
+    second = stable_split_assignments(values, seed="task-1")
+
+    assert first == second
+    assert first.count("train") == 22
+    assert first.count("validation") == 3
+    assert first.count("test") == 3
+
+
 def test_generate_standard_records_supports_json_qa_and_stable_variants() -> None:
     previews = [
         {

compute/engines/llama_factory/adapter.py

@@ -265,6 +265,9 @@ def build_command(config: dict[str, Any], llama_factory_home: str = "/app/LLaMA-
     ]
     if dataset_dir:
         command.extend(["--dataset_dir", str(dataset_dir)])
+    eval_dataset = config.get("eval_dataset")
+    if eval_dataset:
+        command.extend(["--eval_dataset", str(eval_dataset), "--do_eval", "true"])
     _optional_arg(config, command, "--cutoff_len", "max_length", "cutoff_len")
     _optional_arg(config, command, "--lr_scheduler_type", "lr_scheduler_type")
     _optional_arg(config, command, "--warmup_ratio", "warmup_ratio")
@@ -273,6 +276,7 @@ def build_command(config: dict[str, Any], llama_factory_home: str = "/app/LLaMA-
     _optional_arg(config, command, "--lora_alpha", "lora_alpha")
     _optional_arg(config, command, "--lora_dropout", "lora_dropout")
     _optional_arg(config, command, "--gradient_accumulation_steps", "gradient_accumulation_steps")
+    if not eval_dataset:
         _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")
@@ -293,4 +297,3 @@ def parse_log_line(line: str) -> dict[str, float] | None:
         if match:
             result[key] = float(match.group(1))
     return result or None
-

compute/tests/test_llama_factory_adapter.py

@@ -0,0 +1,23 @@
+from __future__ import annotations
+
+from compute.engines.llama_factory.adapter import build_command
+
+
+def test_build_command_uses_explicit_validation_dataset_without_resplitting() -> None:
+    result = build_command(
+        {
+            "base_model": "/models/qwen",
+            "dataset": "ygft_dataset_train",
+            "eval_dataset": "ygft_dataset_validation",
+            "dataset_dir": "/datasets/example",
+            "output_dir": "/outputs/example",
+            "val_size": 0.1,
+        }
+    )
+
+    assert result.command[result.command.index("--dataset") + 1] == "ygft_dataset_train"
+    assert result.command[result.command.index("--eval_dataset") + 1] == (
+        "ygft_dataset_validation"
+    )
+    assert "--do_eval" in result.command
+    assert "--val_size" not in result.command