add lightweight model training skill
This commit is contained in:
@@ -0,0 +1,80 @@
|
||||
# ZK SFT 默认训练配方
|
||||
|
||||
当前轻量训练默认复用 `model-iteration` 的 SFT 提交流程,但只使用其中的训练提交能力。
|
||||
|
||||
## 训练脚本
|
||||
|
||||
```text
|
||||
skills/model-iteration/scripts/prepare_and_train_sft.py
|
||||
skills/model-iteration/scripts/resolve_run_ids.sh
|
||||
skills/model-iteration/scripts/submit_sft.sh
|
||||
skills/model-iteration/scripts/submit_cml_eval.sh
|
||||
skills/model-iteration/scripts/sft_train_job.yaml.tpl
|
||||
skills/model-iteration/assets/config.yaml
|
||||
```
|
||||
|
||||
## 默认基模
|
||||
|
||||
```text
|
||||
/mnt/wangsenhao/verl_zk/Qwen3-4B-Instruct-2507
|
||||
```
|
||||
|
||||
`sft_train_job.yaml.tpl` 内固定从该基模启动 SFT。不要从上一轮 `sft_output` 续训,除非用户明确改变训练策略。
|
||||
|
||||
## 默认数据仓库
|
||||
|
||||
```text
|
||||
git@git.n.xiaomi.com:ai-service/ai-planning.git
|
||||
branch: autoresearch-v1
|
||||
```
|
||||
|
||||
训练脚本会扫描:
|
||||
|
||||
```text
|
||||
ai-planning/data/train_set/*/*.jsonl
|
||||
```
|
||||
|
||||
其中 `*_valid.jsonl` 会作为验证数据,其余训练 jsonl 进入训练集。
|
||||
|
||||
## 默认远端工作区
|
||||
|
||||
```text
|
||||
/mnt/wangsenhao/autoresearch-zk-users/<owner>/<run_name>
|
||||
```
|
||||
|
||||
训练产物:
|
||||
|
||||
```text
|
||||
<workspace>/sft_output/
|
||||
<workspace>/sft_output/_SUCCESS
|
||||
```
|
||||
|
||||
## CML 环境
|
||||
|
||||
CloudML CLI 通常位于:
|
||||
|
||||
```bash
|
||||
source ~/.cloudml-cli/.profile
|
||||
```
|
||||
|
||||
该 profile 会设置:
|
||||
|
||||
```bash
|
||||
PATH=$HOME/.cloudml-cli/bin:$PATH
|
||||
CLOUDML_CONFIG=$HOME/.config/cloudml/config.yaml
|
||||
```
|
||||
|
||||
检查命令:
|
||||
|
||||
```bash
|
||||
cml config show
|
||||
cml custom_train describe <JOB_ID>
|
||||
```
|
||||
|
||||
## 常见坑
|
||||
|
||||
- Jupyter POST 请求缺少 `_xsrf`:从 cookie 取 `_xsrf`,请求头带 `X-XSRFToken`。
|
||||
- websocket 证书校验失败:内网证书场景可在客户端禁用证书校验,但不要降低服务端安全配置。
|
||||
- `which cml` 为空:先 `source ~/.cloudml-cli/.profile`。
|
||||
- `resolve_run_ids.sh` 没有 `RUNDIC`:只使用 `SFT_RUNDIC` / `EVAL_RUNDIC`。
|
||||
- git clone 看似卡住:先检查目标目录是否已经完整、是否存在残留进程,不要重复提交训练。
|
||||
Reference in New Issue
Block a user