Alibaba Cloud Model Studio

https://modelstudio.console.alibabacloud.com
Admin
Alibaba Cloud Model Studio is a comprehensive AI platform providing access to Qwen and other...
Tokens:2,114
Snippets:27
Trust Score:10
Update:1 week ago
Show doc for...
Tokens:
Raw
### Install Bailian CLI via coding agent prompt

Source: https://modelstudio.console.alibabacloud.com

Copy this prompt into a coding agent to install the Bailian CLI and log in using the international site. The prompt references the installation instructions at https://modelstudio.alibabacloud.com/cli/install.md.

```text
$Please read: https://modelstudio.alibabacloud.com/cli/install.md and install the Bailian CLI for me according to the instructions. Additionally, log in using the international site.
```

--------------------------------

### DPO training code example

Source: https://modelstudio.console.alibabacloud.com

Python code snippet for direct preference optimization (DPO) using a preferences dataset and the DPOTrainer. It loads chosen/rejected pairs, sets beta to 0.1, trains, and reports a reward margin of 0.31.

```python
pairs = load_dataset("preferences")
chosen, rejected = pairs["train"]
trainer = DPOTrainer(model, ref_model, pairs)
trainer.beta = 0.1 
trainer.train()
reward_margin => 0.31
```

--------------------------------

### CPT training code example

Source: https://modelstudio.console.alibabacloud.com

Python code snippet for continued pre-training (CPT) using a corpus dataset and the Trainer. It runs CLM training, uses a data collator, resumes from checkpoint, and reports perplexity of 8.64.

```python
corpus = load_dataset("cpt-corpus")
run_clm --model qwen3 --dataset corpus
collator = DataCollatorForLanguageModeling
trainer = Trainer(model, corpus)
trainer.train(resume_from_checkpoint=True)
perplexity => 8.64
```

--------------------------------

### SFT training code example

Source: https://modelstudio.console.alibabacloud.com

Python code snippet for supervised fine-tuning (SFT) using a dataset named 'sft' and the SFTTrainer. It loads the dataset, trains with resume, saves the model, and reports an eval loss of 0.417.

```python
dataset = load_dataset("sft")
messages = dataset["train"]
trainer = SFTTrainer(model, dataset)
trainer.train(resume=True)
trainer.save_model("qwen3-sft")
eval_loss => 0.417
```

--------------------------------

### SFT training command

Source: https://modelstudio.console.alibabacloud.com

Command to run supervised fine-tuning (SFT) with the Bailian CLI. The output shows initialization steps and a success message.

```bash
$ bailian train --task sft
```

```text
SFT Training Initialization
Step 1/3: Loading SFT dataset and template...Loading SFT dataset and template...
Step 2/3: Running supervised SFT tuning...Running supervised SFT tuning...
Step 3/3: Evaluating SFT checkpoint...Evaluating SFT checkpoint...
■ Success! SFT checkpoint exported.
```
Tokens:
Raw