Precomputed Qwen/Qwen3.6-27B pooled embeddings of (state, action) pairs from SOTA
mini-SWE-agent trajectories on SWE-bench Verified. These are the cached embeddings used to
train a retrieval / process reward model (PRM) for best-of-N action selection (pick the
candidate action per step that maximizes cos(proj_state(state), proj_action(action))).
Embeddings are frozen 5120-d vectors; the PRM trains two lightweight… See the full description on the dataset page:
https://huggingface.co/datasets/tarsur385/swev-prm-embeddings-qwen36-27b.
