0.992

README.md

@@ -1,28 +1,148 @@
 # AlphaMaster
 
-![](showcase.png)
+基于深度神经网络强化学习的量化因子挖掘中心：从 Parquet / MT5 K 线自动搜索可解释因子公式，支持 Web 端训练、回测与实时信号分析。
 
-*在 AlphaGPT 开源后，社区曾使用开源版本管理了 1.66 亿人民币规模资金。*
+[![License: AGPL v3](https://img.shields.io/badge/License-AGPL%20v3-blue.svg)](LICENSE)
 
-[时隔数月后，项目作者再次受到了人身安全威胁。](https://github.com/imbue-bit/fuck_all_tian)
+![Web 控制台总览](docs/images/00_hero.png)
 
-也许金融真的是一个意外死亡率很高的行业🤔
+仓库地址：[github.com/rosemarycox5334-debug/AlphaMaster](https://github.com/rosemarycox5334-debug/AlphaMaster)
+
+---
+
+## 它做什么
+
+AlphaMaster 把「挖因子」做成一条可操作的流水线：
+
+1. **训练**：用强化学习在特征 + 算子空间里搜索公式，按验证集表现选优  
+2. **回测**：用 `tanh(因子)` 连续仓位在历史行情上模拟交易，看资金曲线与绩效  
+3. **实时分析**：按周期收盘后重算信号，展示方向与把握；方向转折可推飞书提醒  
+
+公式以 token 序列保存（如 `strategies/best_BTCUSDT.json`），可用 StackVM 解释执行，训练 / 回测 / 实时共用同一套信号逻辑。
+
+---
+
+## Web 控制台（推荐入口）
+
+```bash
+pip install -r requirements.txt
+python run_web.py --port 8765
+```
+
+浏览器打开 [http://127.0.0.1:8765](http://127.0.0.1:8765)。界面分三步：
+
+| 步骤 | 作用 |
+|------|------|
+| **01 模型训练** | 选 Parquet、开始 / 重新训练、看曲线与日志、导出策略与检查点 |
+| **02 策略回测** | 选策略 JSON，设手续费 / 滑点，看绩效与资金曲线 |
+| **03 实时分析** | 多数据源监控，收盘后更新信号；可选飞书转折提醒 |
+
+### 模型训练
+
+![训练页](docs/images/01_train.png)
+
+- Parquet 命名：`{品种}_{周期}.parquet`，例如 `BTCUSDT_H1.parquet`、`XAUUSD_H1.parquet`  
+- **开始训练**：有检查点则断点续训  
+- **重新训练**：清除检查点从头搜索；已有更优策略作为分数下限，不会被弱结果覆盖  
+- 展示最优分数、验证分数、训练曲线与最优公式；可选 AI 分析当前训练情况  
+
+### 策略回测
+
+![回测页](docs/images/02_backtest.png)
+
+- 仓位：`position = tanh(factor)`，信号越强仓位越大  
+- 成本：手续费 + 滑点（默认约 0.02% / 0.01%）  
+- 输出：总收益、夏普、索提诺、盈亏比、滚动夏普与资金曲线  
+
+![资金曲线示例](docs/images/04_equity.png)
+
+### 实时分析
+
+![实时分析页](docs/images/03_realtime.png)
+
+- 数据源：MT5 / OKX 等（以界面可用源为准）  
+- **只在当前周期 K 线收盘后**重新判断；未收盘 bar 不参与信号  
+- 卡片展示方向（看涨 / 看跌 / 不确定）与把握程度  
+- 可选飞书 Webhook：仅在方向转折时推送文字提醒  
+
+---
+
+## 项目结构
+
+```
+AlphaMaster/
+├── web/                 # FastAPI Web UI（训练 / 回测 / 实时）
+├── model_core/          # 特征、算子、StackVM、训练引擎、回测评分
+├── data_pipeline/       # Parquet / MT5 K 线加载与对齐
+├── strategy_manager/    # 实盘信号与仓位逻辑（与回测口径一致）
+├── execution/           # MT5 下单接口
+├── backtest_viz/        # 回测引擎与图表
+├── strategies/          # best_{symbol}.json 策略文件
+├── checkpoints/         # 训练检查点
+├── run_web.py           # 启动 Web 控制台
+├── train_file.py        # CLI：从单个 Parquet 训练
+└── requirements.txt
+```
+
+---
+
+## 环境要求
+
+- Python **3.10+**（建议 3.11）  
+- PyTorch、pandas、FastAPI、uvicorn 等（见 `requirements.txt`）  
+- 可选：MetaTrader 5 终端（实时 MT5 源 / 实盘相关脚本）  
+- 复制 `.env.example` 为 `.env` 填写 MT5 等凭证（`.env` 已 gitignore）  
+
+```bash
+pip install -r requirements.txt
+# 可选可视化等：pip install -r requirements-optional.txt
+```
+
+---
+
+## 常用命令
+
+```bash
+# Web 控制台
+python run_web.py --port 8765
+
+# CLI 训练（自动续训；加 --from-scratch 则重新训练）
+python train_file.py --data-file D:\K线数据\BTCUSDT_H1.parquet
+python train_file.py --data-file D:\K线数据\BTCUSDT_H1.parquet --from-scratch
+```
+
+策略输出默认在 `strategies/best_{symbol}.json`。
+
+---
+
+## 信号口径（训练 / 回测 / 实时一致）
+
+- 因子经 StackVM 算出标量序列  
+- `position = tanh(factor)` ∈ (-1, 1)  
+- `|position|` 小于阈值时视为无信号（观望）  
+- 实时侧只用**已收盘** K 线，避免盘中抖动与回测不一致  
+
+---
+
+## 截图更新
+
+仓库内展示图由当前 Web UI 截取，可用：
+
+```bash
+python scripts/capture_readme_shots.py
+```
+
+（需本机已启动 `python run_web.py --port 8765`，并已安装 Playwright + Chromium。）
 
 ---
 
 ## License
 
-This project is licensed under the [GNU Affero General Public License v3.0 (AGPL-3.0)](LICENSE).
-Derivative works and network deployments must provide corresponding source code under the same license.
-
----
-
-如果您觉得 AlphaGPT 很有用，可以通过以下渠道捐助作者：
-
-![](helpme.jpg)
+本项目采用 [GNU Affero General Public License v3.0 (AGPL-3.0)](LICENSE)。  
+修改、分发或通过网络提供服务时，须按相同协议公开对应源代码。
 
 ---
 
 ## Star History
 
-[![Star History Chart](https://api.star-history.com/svg?repos=imbue-bit/AlphaGPT&type=date&legend=top-left)](https://www.star-history.com/#imbue-bit/AlphaGPT&type=date&legend=top-left)
+[![Star History Chart](https://api.star-history.com/svg?repos=rosemarycox5334-debug/AlphaMaster&type=date&legend=top-left)](https://www.star-history.com/#rosemarycox5334-debug/AlphaMaster&type=date&legend=top-left)

scripts/capture_readme_shots.py

@@ -0,0 +1,25 @@
+from pathlib import Path
+from playwright.sync_api import sync_playwright
+
+out = Path(__file__).resolve().parents[1] / "docs" / "images"
+out.mkdir(parents=True, exist_ok=True)
+url = "http://127.0.0.1:8765/"
+
+with sync_playwright() as p:
+    browser = p.chromium.launch(headless=True)
+    page = browser.new_page(viewport={"width": 1440, "height": 900}, device_scale_factor=1.25)
+    page.goto(url, wait_until="networkidle", timeout=60000)
+    page.wait_for_timeout(2000)
+    page.screenshot(path=str(out / "00_hero.png"), full_page=False)
+    page.screenshot(path=str(out / "01_train.png"), full_page=True)
+
+    page.locator('button[data-page="backtest"]').click()
+    page.wait_for_timeout(1000)
+    page.screenshot(path=str(out / "02_backtest.png"), full_page=True)
+
+    page.locator('button[data-page="realtime"]').click()
+    page.wait_for_timeout(1500)
+    page.screenshot(path=str(out / "03_realtime.png"), full_page=True)
+    browser.close()
+
+print("wrote", sorted(p.name for p in out.glob("*.png")))

strategies/best_BTCUSDT.json

@@ -4,14 +4,12 @@
   "formula": [
     39,
     89,
+    97,
     4,
     87,
-    21,
-    115,
-    87,
-    115
+    31,
+    126,
+    121
   ],
-  "best_score": 1.3544840812683105,
-  "formula_decoded": "YANG_ZHANG_VOL → TS_MAX_10 → SLOPE20 → MOMENTUM_5 → BOLL_POS → MIN → MOMENTUM_5 → MIN",
-  "train_step": 0
+  "best_score": 1.4946367740631104
 }

train_file.py

@@ -28,7 +28,7 @@ from model_core.engine import AlphaEngine
 from model_core.vocab import VOCAB_VERSION
 
 
-def train_from_file(data_file: str) -> AlphaEngine | None:
+def train_from_file(data_file: str, *, from_scratch: bool = False) -> AlphaEngine | None:
     info = inspect_parquet_file(data_file)
     symbol = info["symbol"]
     timeframe = info["timeframe"]
@@ -40,6 +40,7 @@ def train_from_file(data_file: str) -> AlphaEngine | None:
     print(f"  周期: {timeframe}")
     print(f"  训练步数: {ModelConfig.TRAIN_STEPS}")
     print(f"  K线数: {info['bars']}")
+    print(f"  模式: {'重新训练（从头）' if from_scratch else '自动续训'}")
     print(f"{'='*60}")
 
     try:
@@ -57,7 +58,25 @@ def train_from_file(data_file: str) -> AlphaEngine | None:
     ckpt_files = sorted(_glob.glob(ckpt_pattern))
     start_step = 0
 
-    if ckpt_files:
+    if from_scratch:
+        removed = 0
+        for p in ckpt_files:
+            try:
+                pathlib.Path(p).unlink(missing_ok=True)
+                removed += 1
+            except OSError as e:
+                print(f"  [警告] 无法删除检查点 {p}: {e}")
+        hist_path = pathlib.Path(f"training_history_{symbol}.json")
+        if hist_path.exists():
+            try:
+                hist_path.unlink()
+            except OSError:
+                pass
+        print(f"  [重新训练] 已清除 {removed} 个检查点，从第 0 步开始")
+        # 保留已有最优策略作为分数下限，避免开局弱公式覆盖 strategies/best_*.json
+        _seed_best_from_strategy(engine, symbol)
+        ckpt_files = []
+    elif ckpt_files:
         latest = ckpt_files[-1]
         try:
             start_step = engine.load_checkpoint(latest)
@@ -70,7 +89,7 @@ def train_from_file(data_file: str) -> AlphaEngine | None:
         _save_strategy(engine, symbol, timeframe, data_file)
         return engine
 
-    if start_step == 0:
+    if start_step == 0 and not from_scratch:
         hist_path = pathlib.Path(f"training_history_{symbol}.json")
         if hist_path.exists():
             hist_path.unlink()
@@ -84,9 +103,44 @@ def train_from_file(data_file: str) -> AlphaEngine | None:
     return engine
 
 
+def _seed_best_from_strategy(engine: AlphaEngine, symbol: str) -> None:
+    """把已有 best_{symbol}.json 当作重新训练的分数下限。"""
+    path = pathlib.Path("strategies") / f"best_{symbol}.json"
+    if not path.exists():
+        return
+    try:
+        data = json.loads(path.read_text(encoding="utf-8"))
+    except (json.JSONDecodeError, OSError) as e:
+        print(f"  [警告] 读取已有策略失败: {e}")
+        return
+    formula = data.get("formula")
+    score = data.get("best_score")
+    if not formula or score is None:
+        return
+    try:
+        engine.best_formula = [int(t) for t in formula]
+        engine.best_score = float(score)
+        print(f"  [重新训练] 保留已有最优分数下限={engine.best_score:.4f}，仅更好时才会覆盖策略文件")
+    except (TypeError, ValueError) as e:
+        print(f"  [警告] 已有策略无法用作下限: {e}")
+
+
 def _save_strategy(engine: AlphaEngine, symbol: str, timeframe: str, data_file: str) -> None:
     path = pathlib.Path("strategies") / f"best_{symbol}.json"
     path.parent.mkdir(exist_ok=True)
+    # 若磁盘上已有更高分，不要用更弱结果覆盖
+    if path.exists() and engine.best_formula is not None:
+        try:
+            old = json.loads(path.read_text(encoding="utf-8"))
+            old_score = old.get("best_score")
+            if old_score is not None and float(old_score) > float(engine.best_score):
+                print(
+                    f"  [策略] 保留磁盘更优结果 {float(old_score):.4f} "
+                    f"> 本次 {float(engine.best_score):.4f}，未覆盖 {path}"
+                )
+                return
+        except (json.JSONDecodeError, OSError, TypeError, ValueError):
+            pass
     data = {
         "vocab_version": VOCAB_VERSION,
         "symbol": symbol,
@@ -108,7 +162,7 @@ if __name__ == "__main__":
     ModelConfig.REWARD_MODE = "ftmo"
 
     if "--data-file" not in sys.argv:
-        print("用法: python train_file.py --data-file PATH\\TO\\SYMBOL_TF.parquet")
+        print("用法: python train_file.py --data-file PATH\\TO\\SYMBOL_TF.parquet [--from-scratch]")
         print("示例: python train_file.py --data-file D:\\K线数据\\AAPL_H1.parquet")
         sys.exit(1)
 
@@ -118,8 +172,9 @@ if __name__ == "__main__":
         sys.exit(1)
 
     data_file = sys.argv[idx + 1]
+    from_scratch = "--from-scratch" in sys.argv
     t0 = time.time()
-    eng = train_from_file(data_file)
+    eng = train_from_file(data_file, from_scratch=from_scratch)
     elapsed = time.time() - t0
 
     if eng:

web/app.py

@@ -67,6 +67,7 @@ app.add_middleware(
 
 class StartTrainingRequest(BaseModel):
     data_file: str
+    from_scratch: bool = False
 
 
 class ClientLogRequest(BaseModel):
@@ -467,12 +468,21 @@ def _progress_with_live_step(symbol: str, active: bool) -> dict[str, Any]:
             current_step = max(current_step, live)
     train_steps = p.train_steps
     progress_pct = min(100.0, 100.0 * current_step / train_steps) if train_steps > 0 else 0.0
+    val_score = None
+    hist = p.history or {}
+    vals = hist.get("val_score") or []
+    if vals:
+        try:
+            val_score = float(vals[-1])
+        except (TypeError, ValueError):
+            val_score = None
     return {
         "symbol": p.symbol,
         "current_step": current_step,
         "train_steps": train_steps,
         "progress_pct": round(progress_pct, 1),
         "best_score": p.best_score,
+        "val_score": val_score,
         "formula_decoded": p.formula_decoded,
         "status": p.status,
         "history": p.history,
@@ -499,6 +509,7 @@ def api_overview() -> dict[str, Any]:
                 "train_steps": row["train_steps"],
                 "progress_pct": row["progress_pct"],
                 "best_score": row["best_score"],
+                "val_score": row.get("val_score"),
                 "formula_decoded": row["formula_decoded"],
                 "has_checkpoint": row.get("has_checkpoint", False),
                 "has_strategy": row.get("has_strategy", False),
@@ -517,6 +528,7 @@ def api_overview() -> dict[str, Any]:
             "train_steps": row["train_steps"],
             "progress_pct": row["progress_pct"],
             "best_score": row["best_score"],
+            "val_score": row.get("val_score"),
             "formula_decoded": row["formula_decoded"],
             "has_checkpoint": row.get("has_checkpoint", False),
             "has_strategy": row.get("has_strategy", False),
@@ -637,10 +649,18 @@ def api_training_start(req: StartTrainingRequest) -> dict[str, Any]:
             symbol=info["symbol"],
             timeframe=info["timeframe"],
             mode="ftmo",
+            from_scratch=bool(req.from_scratch),
         )
     except RuntimeError as e:
         raise HTTPException(409, str(e)) from e
-    return {"ok": True, "job": job.to_dict(), "data_file": info}
+    if req.from_scratch:
+        invalidate_checkpoint_cache()
+    return {
+        "ok": True,
+        "job": job.to_dict(),
+        "data_file": info,
+        "from_scratch": bool(req.from_scratch),
+    }
 
 
 @app.post("/api/training/stop")

web/static/app.js

@@ -176,6 +176,7 @@ function renderDataFileCard(info) {
     selectedDataFile = null;
     selectedSymbol = null;
     startBtn.disabled = true;
+    if ($("retrainBtn")) $("retrainBtn").disabled = true;
     if ($("exportBtn")) $("exportBtn").disabled = true;
     if ($("exportTrainingBtn")) $("exportTrainingBtn").disabled = true;
     if ($("importTrainingBtn")) $("importTrainingBtn").disabled = true;
@@ -192,22 +193,29 @@ function renderDataFileCard(info) {
       <div class="data-file-path">${info.data_file}</div>
     `;
     startBtn.disabled = true;
+    if ($("retrainBtn")) $("retrainBtn").disabled = true;
     if ($("exportTrainingBtn")) $("exportTrainingBtn").disabled = true;
     if ($("importTrainingBtn")) $("importTrainingBtn").disabled = true;
     return;
   }
 
   card.className = "data-file-card valid";
+  const yearsText = info.years_h1 != null ? `${info.years_h1} 年` : "—";
   card.innerHTML = `
     <div class="data-file-row">
       <div class="item"><span class="label">品种</span><span class="value sym">${info.symbol}</span></div>
       <div class="item"><span class="label">周期</span><span class="value">${info.timeframe}</span></div>
       <div class="item"><span class="label">K线</span><span class="value">${info.bars?.toLocaleString()}</span></div>
+      <div class="item"><span class="label">数据年限</span><span class="value">${yearsText}</span></div>
       <div class="item"><span class="label">进度</span><span class="value" id="fileProgressPct">—</span></div>
-      <span class="path" title="${info.data_file}">${info.filename || info.data_file}</span>
+      <div class="item"><span class="label">已训练</span><span class="value" id="fileElapsedTime">—</span></div>
+      <div class="item"><span class="label">最优分数</span><span class="value score-best" id="fileBestScore">—</span></div>
+      <div class="item"><span class="label">验证分数</span><span class="value score-val" id="fileValScore">—</span></div>
     </div>
+    <div class="path" title="${info.data_file}">${info.filename || info.data_file}</div>
   `;
   startBtn.disabled = false;
+  if ($("retrainBtn")) $("retrainBtn").disabled = false;
 }
 
 function updateBtStartBtn() {
@@ -248,21 +256,50 @@ function renderStrategyFileCard(info) {
   selectedStrategyFile = info.strategy_file;
   selectedStrategySymbol = info.symbol || null;
   card.className = "data-file-card valid";
+  const timeframeItem = info.timeframe
+    ? `<div class="item"><span class="label">周期</span><span class="value">${info.timeframe}</span></div>`
+    : "";
   card.innerHTML = `
     <div class="data-file-row">
       <div class="item"><span class="label">品种</span><span class="value sym">${info.symbol || "—"}</span></div>
-      <div class="item"><span class="label">分数</span><span class="value">${formatScore(info.best_score)}</span></div>
-      <div class="item"><span class="label">版本</span><span class="value">${info.vocab_version || "—"}</span></div>
-      <span class="path" title="${info.strategy_file}">${info.filename || info.strategy_file}</span>
+      ${timeframeItem}
+      <div class="item"><span class="label">最优分数</span><span class="value score-best">${formatScore(info.best_score)}</span></div>
+      <div class="item"><span class="label">词表版本</span><span class="value">${info.vocab_version || "—"}</span></div>
+      <div class="item"><span class="label">公式长度</span><span class="value">${info.formula_decoded ? info.formula_decoded.split("→").length : "—"}</span></div>
     </div>
+    <div class="path" title="${info.strategy_file}">${info.filename || info.strategy_file}</div>
   `;
   updateBtStartBtn();
 }
 
+function formatElapsed(startedAtIso) {
+  if (!startedAtIso) return "—";
+  const started = new Date(startedAtIso).getTime();
+  if (Number.isNaN(started)) return "—";
+  const secs = Math.max(0, Math.floor((Date.now() - started) / 1000));
+  const h = Math.floor(secs / 3600);
+  const m = Math.floor((secs % 3600) / 60);
+  if (h > 0) return `${h}小时${m}分`;
+  return `${m}分钟`;
+}
+
 function updateFileProgress(progress) {
   const el = document.getElementById("fileProgressPct");
-  if (!el || !progress) return;
-  el.textContent = `${progress.current_step} / ${progress.train_steps} (${progress.progress_pct}%)`;
+  if (el && progress) {
+    el.textContent = `${progress.current_step} / ${progress.train_steps} (${progress.progress_pct}%)`;
+  }
+  const bestEl = document.getElementById("fileBestScore");
+  if (bestEl) {
+    bestEl.textContent = progress ? formatScore(progress.best_score) : "—";
+  }
+  const valEl = document.getElementById("fileValScore");
+  if (valEl) {
+    let val = progress?.val_score;
+    if (val == null && progress?.history?.val_score?.length) {
+      val = progress.history.val_score[progress.history.val_score.length - 1];
+    }
+    valEl.textContent = progress ? formatScore(val) : "—";
+  }
 }
 
 const CHART_SERIES = [
@@ -465,14 +502,19 @@ function updateTrainingUI(training) {
   const active = training?.active;
   const pill = $("jobPill");
   const startBtn = $("startBtn");
+  const retrainBtn = $("retrainBtn");
   const stopBtn = $("stopBtn");
 
+  const elapsedEl = $("fileElapsedTime");
+
   if (!job || job.state === "idle") {
     pill.innerHTML = '<i class="pill-dot"></i>空闲';
     pill.className = "pill";
     startBtn.disabled = !selectedDataFile;
+    if (retrainBtn) retrainBtn.disabled = !selectedDataFile;
     stopBtn.disabled = true;
     $("logHint").textContent = "—";
+    if (elapsedEl) elapsedEl.textContent = "—";
     return;
   }
 
@@ -488,8 +530,13 @@ function updateTrainingUI(training) {
   pill.className = "pill " + (job.state === "running" ? "running" : job.state);
 
   startBtn.disabled = active;
+  if (retrainBtn) retrainBtn.disabled = active;
   stopBtn.disabled = !active;
   $("logHint").textContent = job.log_path || "—";
+  if (elapsedEl) {
+    const elapsed = formatElapsed(job.started_at);
+    elapsedEl.textContent = active || elapsed === "—" ? elapsed : `${elapsed}（已停）`;
+  }
 
   const logView = $("logView");
   const atBottom = isViewAtBottom(logView);
@@ -798,7 +845,32 @@ async function startTraining() {
     const res = await fetchJSON("/api/training/start", {
       method: "POST",
       headers: { "Content-Type": "application/json" },
-      body: JSON.stringify({ data_file: selectedDataFile }),
+      body: JSON.stringify({ data_file: selectedDataFile, from_scratch: false }),
+    });
+    selectedSymbol = res.data_file?.symbol || res.job?.symbol;
+    renderDataFileCard(res.data_file);
+    await refreshOverview();
+  } catch (e) {
+    $("debugView").scrollIntoView({ behavior: "smooth", block: "nearest" });
+  }
+}
+
+async function retrainFromScratch() {
+  if (!selectedDataFile) {
+    await logClientError("请先选择数据文件");
+    return;
+  }
+  const ok = window.confirm(
+    "重新训练会清除该品种的检查点，从第 0 步重新搜索。\n" +
+      "已有的更优策略会保留，只有挖到更高分才会覆盖。\n\n" +
+      "确定要重新训练吗？"
+  );
+  if (!ok) return;
+  try {
+    const res = await fetchJSON("/api/training/start", {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ data_file: selectedDataFile, from_scratch: true }),
     });
     selectedSymbol = res.data_file?.symbol || res.job?.symbol;
     renderDataFileCard(res.data_file);
@@ -2176,6 +2248,7 @@ async function init() {
   }
   $("browseBtn").addEventListener("click", browseDataFile);
   $("startBtn").addEventListener("click", startTraining);
+  if ($("retrainBtn")) $("retrainBtn").addEventListener("click", retrainFromScratch);
   $("stopBtn").addEventListener("click", stopTraining);
   $("exportBtn").addEventListener("click", exportStrategy);
   $("exportTrainingBtn").addEventListener("click", exportTraining);

web/static/index.html

@@ -7,7 +7,7 @@
   <link rel="preconnect" href="https://fonts.googleapis.com" />
   <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
   <link href="https://fonts.googleapis.com/css2?family=DM+Sans:ital,opsz,wght@0,9..40,400;0,9..40,500;0,9..40,600;0,9..40,700;1,9..40,400&family=JetBrains+Mono:wght@400;500;600&family=Orbitron:wght@500;600;700;800&display=swap" rel="stylesheet" />
-  <link rel="stylesheet" href="/static/style.css?v=23" />
+  <link rel="stylesheet" href="/static/style.css?v=27" />
   <script src="https://cdn.jsdelivr.net/npm/chart.js@4.4.1/dist/chart.umd.min.js"></script>
 </head>
 <body>
@@ -74,20 +74,22 @@
       <span class="hud-corner bl"></span><span class="hud-corner br"></span>
       <h2>启动训练</h2>
       <div class="launch-form launch-form-file">
-        <div class="launch-row">
-          <button type="button" class="btn btn-secondary" id="browseBtn">选择数据文件</button>
+        <div class="launch-row launch-row-main">
+          <button type="button" class="btn btn-secondary btn-browse" id="browseBtn">选择数据文件</button>
           <div class="data-file-card" id="dataFileCard">
-            <div class="data-file-empty">尚未选择</div>
-          </div>
-          <div class="launch-actions">
-            <button class="btn btn-primary" id="startBtn" disabled>开始训练</button>
-            <button class="btn btn-danger" id="stopBtn" disabled>停止</button>
-            <button class="btn btn-secondary" id="exportBtn" disabled title="下载当前品种的最优策略 JSON">导出策略</button>
-            <button class="btn btn-secondary" id="exportTrainingBtn" disabled title="打包 checkpoint、训练曲线与策略为 zip">导出训练</button>
-            <button class="btn btn-secondary" id="importTrainingBtn" disabled title="上传 .zip 或 .pt，下次训练断点续训">导入训练</button>
-            <input type="file" id="importTrainingFile" accept=".zip,.pt" hidden />
+            <div class="data-file-empty">尚未选择数据文件</div>
           </div>
         </div>
+        <div class="launch-actions" role="group" aria-label="训练操作">
+          <button class="btn btn-primary" id="startBtn" disabled>开始训练</button>
+          <button class="btn btn-secondary" id="retrainBtn" disabled title="清除检查点，从头重新搜索（保留已有更优策略）">重新训练</button>
+          <button class="btn btn-danger" id="stopBtn" disabled>停止</button>
+          <span class="action-sep" aria-hidden="true"></span>
+          <button class="btn btn-secondary" id="exportBtn" disabled title="下载当前品种的最优策略 JSON">导出策略</button>
+          <button class="btn btn-secondary" id="exportTrainingBtn" disabled title="打包 checkpoint、训练曲线与策略为 zip">导出训练</button>
+          <button class="btn btn-secondary" id="importTrainingBtn" disabled title="上传 .zip 或 .pt，下次训练断点续训">导入训练</button>
+          <input type="file" id="importTrainingFile" accept=".zip,.pt" hidden />
+        </div>
         <small class="field-hint">Parquet 命名：<code>{品种}_{周期}.parquet</code>，如 <code>XAUUSD_H1.parquet</code></small>
       </div>
     </section>
@@ -139,6 +141,46 @@
       </section>
     </div>
 
+    <section class="panel howto-panel">
+      <span class="hud-corner tl"></span><span class="hud-corner tr"></span>
+      <span class="hud-corner bl"></span><span class="hud-corner br"></span>
+      <details class="chart-metrics-help bt-howto" open>
+        <summary>训练说明：并行？最优分数 / 验证分数是什么？</summary>
+        <ol class="bt-howto-list">
+          <li>
+            <strong>单品种单周期就已经把 CPU 吃满了</strong>
+            <p>挖因子这一步计算很重：特征、公式搜索、回测验证会把机器算力用到接近满载。这时再并行开第二个品种或另一个周期，大家只是在抢同一批 CPU，谁也跑不快。</p>
+          </li>
+          <li>
+            <strong>并行并不等于整体更快</strong>
+            <p>两路一起跑，往往各自都变慢，总完工时间通常不比「先训完一个再训下一个」更短，有时还更拖。所以默认一次只训一路，把算力集中用在当前任务上更划算。</p>
+          </li>
+          <li>
+            <strong>只盯一个品种，更好观察训练情况</strong>
+            <p>曲线、日志、最优公式都只对应这一组数据，涨没涨分一眼能看清。多品种并行时页面和日志会搅在一起，反而不容易判断「这个品种还值不值得继续挖」。</p>
+          </li>
+          <li>
+            <strong>最优分数是什么</strong>
+            <p>到目前为止挖到的<strong>最好那条公式</strong>的得分。它一般只升不降：一旦发现更好的公式，这个数就会抬一截；长时间不动，说明这段时间还没挖到更强的。</p>
+          </li>
+          <li>
+            <strong>验证分数是什么</strong>
+            <p>当前这一步在<strong>验证集</strong>上算出来的平均表现，反映「最近这批公式整体好不好」。它会上下波动很正常；更值得看的是整体有没有慢慢抬高，而不是盯着单次跳动。</p>
+          </li>
+          <li>
+            <strong>为什么验证曲线会一截一截地掉下去再爬起来？</strong>
+            <p>图上的验证分数不是「当前最优那一条」的得分，而是<strong>这一步随机抽出来的一整批公式的平均分</strong>。挖着挖着，模型会越来越爱重复同一类公式（探索变少），系统就会主动「搅一搅」——加大随机、换一批新配方继续搜。刚搅完时，新配方里好坏掺半，平均值会突然掉一截；随后又慢慢学到更好的组合，曲线再往上爬。所以会看到锯齿状起伏。</p>
+            <p>绿色的最优分数只记「历史最好的那一次」，所以不会跟着掉下去，只会阶梯往上抬。锯齿本身多半是在重新探索，不等于训练坏了。</p>
+          </li>
+          <li>
+            <strong>重新训练有机会找出更好的量化因子</strong>
+            <p>「开始训练」会接着上次的检查点继续挖；「重新训练」会清掉检查点，从零换一条随机路径再搜一遍。因为搜索带随机性，有时能挖到比现在更高的分数。</p>
+            <p>已有的最优策略会当作分数下限留着——只有找到更好的才会覆盖，不会一开局就被弱公式冲掉。代价是要再花一整轮算力，适合本轮训完或最优分数长期不动时再用。</p>
+          </li>
+        </ol>
+      </details>
+    </section>
+
     <section class="panel strategies-panel">
       <span class="hud-corner tl"></span><span class="hud-corner tr"></span>
       <span class="hud-corner bl"></span><span class="hud-corner br"></span>
@@ -217,8 +259,8 @@
         <span class="hud-corner bl"></span><span class="hud-corner br"></span>
         <h2>启动回测</h2>
         <div class="launch-form launch-form-file">
-          <div class="launch-row">
-            <button type="button" class="btn btn-secondary" id="btBrowseStrategyBtn">选择策略</button>
+          <div class="launch-row launch-row-main">
+            <button type="button" class="btn btn-secondary btn-browse" id="btBrowseStrategyBtn">选择策略</button>
             <div class="data-file-card" id="btStrategyCard">
               <div class="data-file-empty">尚未选择策略文件</div>
             </div>
@@ -500,6 +542,6 @@
   </div>
 
   <script src="/static/bg.js?v=1"></script>
-  <script src="/static/app.js?v=52"></script>
+  <script src="/static/app.js?v=55"></script>
 </body>
 </html>

web/static/style.css

@@ -438,6 +438,7 @@ select:focus { border-color: var(--accent); }
 .launch-actions {
   display: flex;
   flex-wrap: wrap;
+  align-items: center;
   gap: 10px;
   margin-left: auto;
 }
@@ -547,6 +548,33 @@ select:focus { border-color: var(--accent); }
 
 .launch-row .launch-actions { margin-left: auto; }
 
+/* 主行：选择按钮与信息卡等高并排 */
+.launch-row-main { align-items: stretch; }
+
+.btn-browse {
+  flex-shrink: 0;
+  display: inline-flex;
+  align-items: center;
+  justify-content: center;
+  white-space: nowrap;
+}
+
+/* 操作按钮独占一行（训练页）：不再右推，左对齐成组 */
+.launch-form-file > .launch-actions {
+  margin-left: 0;
+  margin-top: 2px;
+}
+
+/* 控制组 / 文件操作组之间的竖线分隔 */
+.action-sep {
+  align-self: stretch;
+  width: 1px;
+  min-height: 22px;
+  margin: 0 5px;
+  background: var(--border-strong);
+  opacity: 0.7;
+}
+
 .bt-cost-row {
   align-items: flex-end;
   gap: 14px;
@@ -600,10 +628,15 @@ select:focus { border-color: var(--accent); }
 .data-file-card {
   flex: 1 1 280px;
   min-width: 0;
+  display: flex;
+  flex-direction: column;
+  justify-content: center;
+  gap: 4px;
+  min-height: 50px;
   background: rgba(14, 20, 31, 0.6);
   border: 1px dashed var(--border-strong);
   border-radius: 10px;
-  padding: 9px 13px;
+  padding: 7px 16px;
   transition: border-color 0.2s, box-shadow 0.2s;
 }
 
@@ -617,34 +650,53 @@ select:focus { border-color: var(--accent); }
 
 .data-file-empty { color: var(--text-muted); font-size: 0.85rem; }
 
+/* 信息卡：仪表盘统计条（每个指标竖向小格 + 细竖线分隔）*/
 .data-file-row {
   display: flex;
   align-items: center;
-  gap: 16px;
+  gap: 0;
   flex-wrap: wrap;
   font-size: 0.85rem;
 }
 
-.data-file-row .item { display: flex; align-items: baseline; gap: 6px; white-space: nowrap; }
+.data-file-row .item {
+  display: flex;
+  flex-direction: column;
+  gap: 3px;
+  padding: 2px 13px;
+  border-left: 1px solid var(--border);
+  white-space: nowrap;
+}
+.data-file-row .item:first-child { border-left: none; padding-left: 2px; }
 
 .data-file-row .label {
-  font-size: 0.68rem;
+  font-size: 0.6rem;
   color: var(--text-muted);
   text-transform: uppercase;
-  letter-spacing: 0.04em;
+  letter-spacing: 0.08em;
 }
 
-.data-file-row .value { font-family: var(--mono); color: var(--text); }
-.data-file-row .value.sym { font-weight: 600; color: var(--accent); text-shadow: 0 0 12px var(--accent-glow); }
-
-.data-file-row .path {
+.data-file-row .value {
   font-family: var(--mono);
-  font-size: 0.72rem;
+  font-size: 0.95rem;
+  line-height: 1.1;
+  color: var(--text);
+}
+.data-file-row .value.sym { font-weight: 600; color: var(--accent); text-shadow: 0 0 12px var(--accent-glow); }
+.data-file-row .value.score-best { color: #34f5c8; font-weight: 600; }
+.data-file-row .value.score-val { color: #38bdf8; font-weight: 600; }
+
+/* 文件名：独立一行放在指标条下方，带顶部细分割线，避免右侧留白 */
+.data-file-card > .path {
+  margin-top: 4px;
+  padding-top: 5px;
+  border-top: 1px solid var(--border);
+  font-family: var(--mono);
+  font-size: 0.7rem;
   color: var(--text-muted);
   overflow: hidden;
   text-overflow: ellipsis;
   white-space: nowrap;
-  max-width: 200px;
 }
 
 .data-file-error { color: var(--danger); font-size: 0.82rem; }

web/training_manager.py

@@ -82,6 +82,8 @@ class TrainingManager:
         symbol: str,
         timeframe: str,
         mode: str = "ftmo",
+        *,
+        from_scratch: bool = False,
     ) -> TrainingJob:
         with self._lock:
             self._refresh_state()
@@ -106,6 +108,8 @@ class TrainingManager:
                 "--data-file",
                 data_file,
             ]
+            if from_scratch:
+                cmd.append("--from-scratch")
 
             self._log_fp = open(log_path, "w", encoding="utf-8", buffering=1)
             env = os.environ.copy()