sft-launch-jupiter

You can launch SFT experiments on the Jupiter cluster using the LLaMA-Factory backend via hpc.launch. This workflow manages multi-node configurations, dataset mixing, and sbatch submission to run end-to-end training with minimal manual setup.

The consolidate and upload flow handles large 32B model configurations by running training with no-upload, followed by checkpoint consolidation and uploading. This process ensures proper checkpoint management and postprocess steps for end-to-end training workflows.

Yes, the workflow supports dataset mixing and concatenation for vanilla Qwen3-8B, 32B, and Qwen3.5 variants. It provides guidance for custom CLI flags and dataset handling within the hpc.launch orchestration on the Jupiter cluster.

Yes, submitting sbatch jobs requires access to the Jupiter cluster and proper conda or environment activation. You also need configuration files under sft/lf_configs to properly define and launch the training workflows.

You can configure multi-node and multi-GPU settings for Qwen3-8B training using configuration files under sft/lf_configs. The hpc.launch backend orchestrates these setups, handling preambles and custom CLI flags for the sbatch submission.