Anyscale Fine-tuning

Fine-tune open-source models on Anyscale's managed infrastructure using Ray for distributed training. Leverage DeepSpeed and FSDP for efficient, scalable fine-tuning. This action pack guides you through the process.

intermediate60 min5 steps

The play

Set up Anyscale CLI
Install and configure the Anyscale CLI to interact with the Anyscale platform. Ensure you have an Anyscale account and API key.
Create a Fine-tuning Job Specification
Define a YAML file specifying the model, dataset, and hyperparameter configurations for your fine-tuning job. Refer to Anyscale documentation for the schema.
Submit the Fine-tuning Job
Use the Anyscale CLI to submit the fine-tuning job, referencing the YAML specification file.
Monitor the Job
Track the progress of your fine-tuning job using the Anyscale dashboard or CLI. Monitor metrics such as loss and training time.
Deploy the Fine-tuned Model
Once the job completes successfully, deploy the fine-tuned model for inference. Follow Anyscale's deployment guidelines for optimal performance.

Starter code

model_id: meta-llama/Llama-2-7b-hf
dataset_id: your_s3_bucket/your_dataset.jsonl
num_epochs: 3
lr: 2e-5

Source

Articleanyscale.com