From b3f4847683c602b7a938b7db13e60a6000b6ee1d Mon Sep 17 00:00:00 2001
From: bingshuo <565183519@qq.com>
Date: Mon, 18 May 2026 11:07:18 +0800
Subject: [PATCH] =?UTF-8?q?=E9=93=B8=E6=B8=8A:=20=E6=B7=BB=E5=8A=A0?=
 =?UTF-8?q?=E4=B9=8B=E4=B9=8Bv2=E8=AE=AD=E7=BB=83=E5=90=AF=E5=8A=A8?=
 =?UTF-8?q?=E9=85=8D=E7=BD=AE=20YAML?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 scripts/training/zhizhi_v2/zhizhi_v2.yaml | 41 +++++++++++++++++++++++
 1 file changed, 41 insertions(+)
 create mode 100644 scripts/training/zhizhi_v2/zhizhi_v2.yaml

diff --git a/scripts/training/zhizhi_v2/zhizhi_v2.yaml b/scripts/training/zhizhi_v2/zhizhi_v2.yaml
new file mode 100644
index 0000000..89030d4
--- /dev/null
+++ b/scripts/training/zhizhi_v2/zhizhi_v2.yaml
@@ -0,0 +1,41 @@
+# 之之母模型 v2.0 · 全参数 SFT
+# 由铸渊配置 · 2026-05-18
+# 基座: Qwen2.5-7B · 数据: 铸渊仓库 + 霜砚 混合语料
+
+### 模型 ###
+model_name_or_path: Qwen/Qwen2.5-7B
+
+### 数据 ###
+dataset: zhizhi_v2
+dataset_dir: /root/autodl-tmp
+template: qwen
+cutoff_len: 4096
+
+### 训练 ###
+finetuning_type: full
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 5.0e-6
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.03
+
+### 精度 ###
+optim: adamw_torch
+bf16: true
+mixed_precision: bf16
+ddp_find_unused_parameters: false
+
+### DeepSpeed ###
+deepspeed: /root/autodl-tmp/ds_config.json
+
+### 输出 ###
+output_dir: /root/autodl-tmp/output/zhizhi_v2
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+preprocessing_num_workers: 16
+
+### 报告 ###
+report_to: none