From 42a4349e69a609d2c4909520cf14b8879c2412a7 Mon Sep 17 00:00:00 2001 From: iomgaa Date: Sat, 18 Jul 2026 08:11:01 -0400 Subject: [PATCH] =?UTF-8?q?=E5=B1=821/T2:=20=E7=94=9F=E6=88=90=E8=84=9A?= =?UTF-8?q?=E6=9C=AC=E6=8E=A5=E5=85=A5=E6=9C=AC=E5=9C=B0=E5=B7=B2=E6=9C=89?= =?UTF-8?q?=E7=9A=84=20DAPO=20=E5=8E=BB=E9=87=8D=E7=89=88=20parquet?= =?UTF-8?q?=EF=BC=8817917=20=E8=A1=8C=EF=BC=89=EF=BC=8C=E8=A1=A5=E8=AF=95?= =?UTF-8?q?=E8=B7=91=E8=AF=B4=E6=98=8E?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 数据管线已用真实数据验证（加载→归一→seed 抽子集，5 条无报错） Co-Authored-By: Claude Fable 5 --- scripts/generate_teacher_completions.py | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/scripts/generate_teacher_completions.py b/scripts/generate_teacher_completions.py index 16447d8..a7fe366 100644 --- a/scripts/generate_teacher_completions.py +++ b/scripts/generate_teacher_completions.py @@ -19,9 +19,14 @@ from ars_opd.teacher import TeacherClient, generate_completions # 非显然约束：这里的 dataset/subset_size/seed 必须与 T5 训练脚本完全一致—— # 两侧各自走"加载→归一→抽子集"，seed 相同才是同一批题（data.py 有详注） -DATASET_PATH = "data/dapo-math-17k.parquet" +DATASET_PATH = "data/dapo-math-17k-unique.parquet" # DAPO 官方去重版，17917 行 CACHE_PATH = "data/teacher_completions_dapo1k_minimax-m3.jsonl" +# 试跑说明：首次建议把下面 subset_size 临时改成 5，跑通并人工抽查缓存里的解答 +# 质量（think 是否剥净、格式是否正常）后再改回 1000 重跑。放心改：subset 是对 +# 同一 seed 的洗牌序列取前缀，前 5 条与前 1000 条的头 5 条完全相同，试跑写入的 +# 缓存在正式跑时全部命中，一分钱不浪费。 + sft_cfg = SFTConfig( dataset_path=DATASET_PATH, output_dir="outputs/_unused", # 本脚本不训练，仅复用数据管线配置