-
Upload DeepSWE-Step198 checkpoint
Training configuration:
- Base model: Qwen/Qwen3-32B
- Training method: RLHF with PPO (verl framework)
- Hardware: 2 nodes x 8 H200 GPUs (16 GPUs total), TP=16 cross-node
- Training started: 2026-02-12 17:35 CST
- Checkpoint: global_step_198 out of 200 total steps
- Batch size: 4 prompts x 8 rollouts = 32 trajectories per step
- PPO epochs: 4
- Dataset: R2E-Gym subset (4578 tasks)
- FSDP parallelism: Ulysses Sequence Parallel = 8
- Checkpoint save frequency: every 2 steps
- Estimated training time: ~67 hours for 200 steps -
initial commit