层1/T2: max_tokens 放大至 16384（防截断，上限不计费）；并发提到 16；进度行加速度与预计剩余时间

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

ars_opd/configs.py

@@ -125,17 +125,17 @@ class TeacherGenConfig:
     top_p: float = 0.95
     """MiniMax M 系官方推荐采样参数：temperature=1.0, top_p=0.95。"""
 
-    max_tokens: int = 8192
+    max_tokens: int = 16384
-    """teacher 单条回复的 token 上限。非显然约束：M3 的思考段也计入此额度，
+    """teacher 单条回复的 token 上限。这是上限不是目标——按实际生成量计费，
-    设太小会把解答挤没（只剩被截断的思考）；student 侧超长解答由 collator 的
+    放大它不增加正常解答的成本，只给最难的题留出写完的空间（8192 时 59 条实测
-    completion 预算兜住，这里宁可给足。"""
+    截断 2 条）。非显然约束：M3 的思考段也计入此额度，设太小会把解答挤没。"""
 
     strip_think: bool = True
     """剥离 content 开头的 <think>...</think> 思考段。SFT 的监督目标是最终
     解答；student 以 enable_thinking=False 训练，学思考段会与模板约定矛盾。"""
 
-    concurrency: int = 8
+    concurrency: int = 16
-    """并发请求数（线程池大小）。"""
+    """并发请求数（线程池大小）。上限看网关的承受力，报 429 就调小。"""
 
     max_retries: int = 3
     """单请求的网络级重试次数（openai 客户端内建指数退避）。"""

ars_opd/teacher.py

@@ -16,6 +16,7 @@ from __future__ import annotations
 import json
 import os
 import re
+import time
 from concurrent.futures import ThreadPoolExecutor, as_completed
 from pathlib import Path
 
@@ -140,6 +141,7 @@ def generate_completions(
 
     failures: list[tuple[str, str]] = []
     finished = 0
+    start = time.monotonic()
     # 写盘收口在主线程（as_completed 消费端），工作线程只跑网络请求——
     # 多线程同写一个文件句柄会交错损坏 JSONL
     with open(path, "a", encoding="utf-8") as f:
@@ -155,7 +157,15 @@ def generate_completions(
                 finally:
                     finished += 1
                     if finished % 20 == 0 or finished == len(todo):
-                        print(f"[teacher] {finished}/{len(todo)} 完成", flush=True)
+                        elapsed = time.monotonic() - start
+                        rate = finished / elapsed * 60  # 条/分
+                        eta = (len(todo) - finished) / rate if rate > 0 else 0
+                        print(
+                            f"[teacher] {finished}/{len(todo)} 完成 | "
+                            f"{rate:.1f} 条/分 | 已用 {elapsed / 60:.1f} 分 | "
+                            f"预计剩余 {eta:.0f} 分",
+                            flush=True,
+                        )
                 record = {
                     "key": key,
                     "completion": completion,