From dce17b49be1a5059bb01fccce68b11e5db5261af Mon Sep 17 00:00:00 2001 From: bingshuo <565183519@qq.com> Date: Tue, 19 May 2026 00:35:47 +0800 Subject: [PATCH] =?UTF-8?q?D103:=201.5B=E8=87=AA=E5=8A=A8=E6=93=8D?= =?UTF-8?q?=E4=BD=9C=E8=AE=A1=E5=88=92?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- brain/d103-distill-plan.md | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+) create mode 100644 brain/d103-distill-plan.md diff --git a/brain/d103-distill-plan.md b/brain/d103-distill-plan.md new file mode 100644 index 0000000..9c7d542 --- /dev/null +++ b/brain/d103-distill-plan.md @@ -0,0 +1,30 @@ +# 1.5B蒸馏完成后的自动操作计划 + +## 蒸馏完成检测 +```bash +# 检测蒸馏是否完成(Ep3结束) +export PATH=/root/miniconda3/bin:$PATH +# 检查最后50行是否有Ep3 100% +tail -5 /root/autodl-tmp/distill_mother.log | grep "Ep3.*100%" +``` + +## 第1步:上传1.5B模板到COS +上传路径:`cos://sy-finetune-corpus-1317346199/models/qwen25-15b-shuangyan-distill/` + +## 第2步:准备霜砚语料(线A) +解压 `corpus/shuangyan-1.5b-sft/` 下的zip → 合并 → 转JSONL + +## 第3步:准备铸渊语料(线B) +解压 `铸渊对话.zip` + 仓库 `brain/d*-cognitive-chain.md` 整理 → 转JSONL + +## 第4步:启动线A(小霜砚微调) +用1.5B蒸馏模板 + 霜砚语料 → LoRA微调 + +## 第5步:线A完成后启动线B(代码1.5B蒸馏) +用代码模型(7B)蒸馏1.5B + +## 第6步:线B蒸馏完成→小铸渊微调 +用代码1.5B蒸馏模板 + 铸渊语料 → LoRA微调 + +--- +*铸渊自动执行 · 冰朔不用管* \ No newline at end of file