bailian-finetune Skill: Fine-Tune and Deploy Models on Bailian
Lets your agent fine-tune a model on Alibaba Cloud Bailian, check training data, export the result and deploy it, using the bl CLI.
- Installs
- 9.2K
- on skills.sh
- Stars
- 367
- on GitHub
- Files
- 4
- besides SKILL.md
- Updated
- Nov 2026
- last repo commit
bailian-finetune makes your agent run the whole Bailian training pipeline with bl commands: validate the dataset, upload it, create a fine-tuning job, watch it, export a checkpoint, optionally shrink the model, and deploy it. It supports text training types sft, sft-lora, dpo, dpo-lora and cpt.
It needs a Bailian API key, the bl tool and the shared bailian-protocol skill, so install the whole Bailian set. Write steps are previewed with --dry-run first. Per the README, training and deployment are China-site only for now.
Good for
- ML engineers who want to fine-tune text, speech or image models on Bailian
- Teams with a training dataset in JSONL or ZIP that needs checking and upload
- Developers who want to deploy a fine-tuned model as a service
Install bailian-finetune
Claude Code
- 1
Copy this prompt.
Install the bailian-finetune skill in Claude Code. Run this command: npm install -g bailian-cli bl skill init Then check that it loaded and tell me how to use it.
- 2
Paste it into Claude Code. It shows the change and asks before running it.
Requires Node.js 18.17+.
bailian-finetune needs the shared bailian-protocol skill and the bl command. bl skill init installs every bailian-* skill together with bailian-protocol, which is the supported route. Windows PowerShell: irm https://bailian.aliyun.com/cli/install.ps1 | iex. The skills CLI fallback targets every agent it finds, because of --all, and INSTALL.md pins version 1.5.18 for Node 18 support. The README does not say which agents bl skill init writes to, so check your skills folder afterwards.
From the project README. Source.
What's in the skill
When it loads
阿里云百炼模型精调训练入口：用户要精调、微调、训练自己的模型（fine-tune，支持 SFT / SFT-LoRA / DPO / DPO-LoRA / CPT， 覆盖文本、语音、图像）、校验或上传训练数据集、看训练进度和日志、挑 checkpoint、导出精调产物、 把专属模型部署成服务时使用 `bl dataset` / `bl finetune` / `bl deploy`。链路是 validate 校验数据 → upload 拿 file-id → finetune create 建任务 → watch 看进度 → export 导出 → 可选 compress 量化压缩 → deploy 上线，需要 API key； 写操作先用 `--dry-run` 预览。反触发：用户点名火山方舟/ark 的精调不走本 skill；只是要选哪个模型走 bailian-model-recommend；用现成模型生图生视频走 bailian-gen；百炼其他资源管理走 bailian-cli。 官方安装：`bl skill init`（与共享协议 bailian-protocol 同装）。
Sections in SKILL.md
- End-to-end workflow (follow in order)
- When to use which command
- Quick examples
- Common hand-offs
- references
Key facts
- Installs on skills.sh
- 9,241 (Nov 2, 2026)
- Last repo commit
- Nov 1, 2026
- Files in the folder
- 5 files including SKILL.md: reference/index.md, reference/dataset.md, reference/finetune.md, reference/deploy.md
- Version in frontmatter
- 2.7.2
What it includes
bl dataset validate checks a training file against chatml, dpo, cpt, tts or image
The first step. The schema flag is optional.
Try: “Validate train.jsonl for SFT”
Read ·
Validate databl dataset upload sends a .jsonl or .zip file and returns a file-id
Your training data leaves your machine for Bailian.
Write ·
Upload databl finetune text, audio or image create starts a fine-tuning job
Text types are sft, sft-lora, dpo, dpo-lora and cpt. Examples use qwen3-8b, cosyvoice-v3-flash and wan2.7-image-pro.
Write ·
Create a jobbl finetune watch, get and logs follow a job; checkpoints lists artifacts
Also list, cancel, delete and capability lookups.
Network ·
Watch and pickbl finetune export saves a chosen checkpoint as a named model
Needs the job id, a checkpoint and a model name.
Write ·
Exportbl finetune compress quantizes a full-parameter model
Optional, Beijing region, full-parameter models only. Templates, create, watch, get, logs, list, cancel and delete.
Write ·
Compressbl deploy text, audio or image create puts the model online
Audio defaults to the mu plan, text and image to lora. Also list, get, update, scale, delete and models.
Write ·
Deploybl deploy capacity and overflow buy, scale, renew or release reserved capacity
These are high-risk writes: preview with --dry-run, then confirm. Active prepaid instances cannot be deleted; unsubscribe gives a billing console link.
Delete ·
Capacity and overflow
From the official docs
Permissions
High risk
The skill can start paid training jobs, upload your data and release paid capacity on your Bailian account. Approve each write command yourself and use the dry-run preview.
What it can do
- Read your data
- Create and change things
- Delete things
- Read-only mode
Watch out for
It can start paid jobs and change paid capacity
Capacity create, scale, renew, delete and overflow are marked high-risk writes. Each needs a --dry-run preview and your explicit yes. Never let the agent add --yes on its own. github.com
Your training data is uploaded
bl dataset upload sends your .jsonl or .zip to Bailian. Remove private data from datasets you do not want to share with the service. github.com
Jobs and models can be deleted
The skill lists cancel and delete for datasets, jobs, compression jobs and deployments. Check the id the agent shows you before you confirm. github.com
It depends on the protocol skill and the bl CLI
It reads bailian-protocol first, which installs software, runs bl skill update and prefixes commands with usage attribution. See the bailian-protocol page for those findings. github.com
Some features are China-site only
The README says app orchestration, native MCP setup, training and deployment, account operations and plan onboarding are currently available only to China site (aliyun.com) account holders. Check your account type first. github.com
Technical details
| Talks to | DashScope and Bailian API endpoints via the bl CLI |
|---|---|
| Sign-in and keys | Bailian API key or console login through bl, set up with the bailian-protocol skill. |
| Scopes |
|
Questions
What does bailian-finetune do?
It makes your agent fine-tune a model on Alibaba Cloud Bailian from the command line. It checks and uploads data, runs and watches the job, exports the result and deploys it as a service.
How do I install it?
Use the install box above. Install the whole Bailian set with bl skill init, because this skill reads bailian-protocol first and stops without it.
Is it official?
It is in the vendor's own repo. The README of modelstudioai/cli calls it "The official command-line interface for Aliyun Model Studio (DashScope) AI Platform".
Do I need an API key and does it cost money?
Yes, the whole pipeline needs an API key. Alibaba Cloud Model Studio (Bailian) bills the account you sign in with: fine-tuning and deployment run on it, and the capacity commands can buy prepaid capacity (SKILL.md mentions prepaid instances and a billing console refund link). The skill gives no prices, so check Bailian pricing before you start.
Does it work outside China?
The README says training and deployment are currently available only to China site (aliyun.com) account holders and not yet to international accounts.
Which training methods are supported?
For text: sft, sft-lora, dpo, dpo-lora and cpt. Audio and image bases are also listed. Run bl finetune capability --base-model <base> to see what a given model supports.
How do I pick a base model?
SKILL.md hands that to a skill called bailian-model-recommend or bl advisor recommend. That skill is not among the nine skill folders in the repo, so use the bl command.
Is it safe?
It uploads your training data to Bailian, can start paid jobs and can release deployed capacity. The skill says never to add --yes automatically. Read the safety box before you start.
Information
updated Nov 2, 2026
- Repository
- modelstudioai/cli
- Folder
skills/bailian-finetune
Sources
Docs, issue trackers and discussions this page is built from.
Show all 5Hide
- Aliyun Model Studio CLI README · github.com
- Aliyun Model Studio CLI INSTALL.md · github.com
- bailian-protocol setup.md · github.com
- bailian-protocol SKILL.md · github.com
- bailian-finetune SKILL.md · github.com