| # Phase‑1 synchronous baseline configuration | |
| # Minimal settings to run a single‑node PPO baseline with colocated rollout & trainer. | |
| # Adjust GPU ids, model path, and data paths as needed. | |
| model: | |
| name: "Qwen2.5-0.5B" | |
| path: "${MODEL_PATH}" # set via env var or replace here | |
| training: | |
| epochs: 3 | |
| batch_size: 64 | |
| learning_rate: 1e-5 | |
| optimizer: "adamw" | |
| loss: "ppo" | |
| clip_eps: 0.2 | |
| rollout: | |
| backend: "vllm" # or "sglang" | |
| max_new_tokens: 256 | |
| temperature: 0.7 | |
| top_p: 0.95 | |
| replay_buffer: | |
| max_size: 10000 | |
| # No disaggregation; trainer and rollout share the same GPUs. | |