The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
Qwen3-30B-A3B SWE-rebench OpenHands router traces
Router activations from nebius/SWE-rebench-openhands-Qwen3-30B-A3B replaying OpenHands coding-agent
trajectories from nebius/SWE-rebench-openhands-trajectories with the model's native chat template.
The source trajectories were collected with Qwen3-Coder-480B-A35B-Instruct; the
30B checkpoint used here was subsequently fine-tuned on that trajectory dataset.
- Run:
qwen3-30b-a3b-swe-rebench-router-traces-2026-07-21-17-22-44 - Context lengths: 512, 1024, 2048, 4096, 8192
- Total tokens: 1007104
- Model revision:
2b25b529d3467ee5fcb2cde0e9711215974012b0 - Dataset revision:
35455389ab51bf5e2306bfd436ef72d0f98bf882 - Attention backend:
flash_attention_2
Each .pt shard is a serialized TensorDict with input_ids, router_logits,
full-expert router_probs, actual top-k topk_indices, normalized topk_probs,
trajectory_id, instance_id, source_token_count, and resolved. Shapes are
documented in each run's manifest.json. Static router matrices are stored once
in router_parameters.pt. Train/eval assignment is grouped by SWE-rebench
instance to prevent trajectories for the same task from crossing the split.
- Downloads last month
- 212