Quick Start

Quick Start#

This page walks the fastest path — environment → model and data → launch → verification — running conf/qwen3_4b/dapo_h800_1node (Qwen3-4B + DAPO) on a single 8-GPU (H800) node to get the first training step going. Other launch options (multi-node, …) and run modes live in the Run Guide; every field is documented in the Config Reference.

1. Prepare the environment#

Since LoongSage may include version-specific builds and temporary patches for SGLang/Megatron, we strongly recommend using the latest Docker image we provide to avoid environment issues.

Pull and start the Docker container#

# pull the latest image
docker pull loongsage/loongsage:latest

docker run --rm --gpus all --ipc=host \
  -it loongsage/loongsage:latest /bin/bash

2. Prepare model and data#

Download the model and dataset#

This preset uses Qwen3-4B. Download the weights straight from Hugging Face:

hf download Qwen/Qwen3-4B --local-dir /root/Qwen3-4B

This preset is a DAPO math task and uses the DAPO-Math-17k training set:

hf download Haitao999/DAPO-Math-17k-unique --repo-type=dataset \
  --local-dir /root/DAPO-Math-17k-unique

3. Launch training#

Launch#

There are two equivalent ways to launch:

Method 1 — fill in the config file, then launch with examples/start.sh (recommended for the first run). Fill the two ??? directly in conf/qwen3_4b/dapo_h800_1node.yaml.

Fill in the required fields

hf_model_path: /root/Qwen3-4B
data_source:
  dataset:
    prompt_data_path: /root/DAPO-Math-17k-unique/<file>.parquet

Run

bash examples/start.sh qwen3_4b/dapo_h800_1node

Method 2 — leave the config untouched and override on the command line. examples/start.sh forwards everything after the config name straight to Hydra:

bash examples/start.sh qwen3_4b/dapo_h800_1node \
  hf_model_path=/root/Qwen3-4B \
  data_source.dataset.prompt_data_path=/root/DAPO-Math-17k-unique

4. Verify the results#

tracking.tracking_backend defaults to console, so metrics go straight to the log. Watch these keys:

Metric

Meaning

rollout/completed_count

Trajectories collected this step; equals num_prompts_per_step × num_trajectories_per_prompt

rollout/reward_mean

Mean reward

train/is_approx_k3_kl

Approximate KL between the current policy and the sampling policy

train/pg_loss, train/entropy

Training loss, entropy

timing/*

Per-phase wall time

5. Next steps#