Contrastive-LM/tb4-clm-train-embeddings-8k
Terminal trajectory embedding task subset 20 selected tasks, 1,281 trajectories, and 123,849 state/action pairs. Only trajectories with numeric reward == 1 are retained. The saved vectors are exact selected rows of the existing embeddings; the encoder was not rerun. train/metadata.json matches both tensor row orders. source_row_indices.json records the original row indices; selection.json records selection parameters, source checksums and output checksums. selected_tasks.json… See the full description on the dataset page: https://huggingface.co/datasets/Contrastive-LM/tb4-clm-train-embeddings-8k.
Terminal trajectory embedding task subset
20 selected tasks, 1,281 trajectories, and 123,849 state/action pairs. Only trajectories with numeric reward == 1 are retained.
The saved vectors are exact selected rows of the existing embeddings; the encoder was not rerun. train/metadata.json matches both tensor row orders. source_row_indices.json records the original row indices; selection.json records selection parameters, source checksums and output checksums.
selected_tasks.json identifies the selected tasks. heldout_tasks.json lists the 46 other tasks, whose vectors remain in the full source. If this subset is used for training, evaluate on held-out tasks, excluding these selected tasks. A trajectory-level success label is not an action correctness label. All source context limits and truncation metadata remain unchanged.
Source embedding dataset: Hkang/terminal-bench-4-qwen3-8b-embeddings, commit 796ecf6ad08e4deaa06507fb5418ca60a74489ba.
This is a reproducible example using seed 42 and successful trajectories only; the full 66-task source remains available. It uses the original Qwen3-8B 8K embeddings with their existing truncation metadata.