Note SFT terminal trajectories | Teacher: Synthetic Terminal-Task-Gen; teacher per subset must be inspected | Terms: CC-BY-4.0; adapter seeds and repository terms separate | Use: SFT + inspect task synthesis recipe | Dataset-adapter versus skill-based configs overlap other SWE/data tasks. A large trace corpus is not automatic release of every underlying verifier/image. Decontam Terminal-Bench seeds.

Note SFT trajectories | Teacher: GLM-4.7-AWQ; Terminus2/Daytona | Terms: Apache-2.0 dataset; SWE-Smith and StackExchange upstream obligations | Use: Broad terminal/SWE SFT | Source mixture includes SWE-Smith, IssueTasks, SuperUser, Tezos. Dedup with original source families; ≥5turn filter is not proof every trace succeeds.

Note RL task dataset | Teacher: Generated task pipeline; optional proprietary-model difficulty filtering | Terms: No explicit HF license metadata; inspect underlying task/project terms | Use: Train RL after license and oracle checks | ~720tasks per card; train files available. Terminal-Bench dev-derived tasks must not contaminate final Terminal-Bench evaluation. Instructions alone insufficient: check images/verifiers actually present.

Note RL task dataset + Harbor environments | Teacher: Five-stage synthesis uses Claude Agent SDK; exact per-artifact teacher check needed | Terms: MIT data; code repo licensing independently unresolved | Use: Train RL, not public evaluation reuse | 602Harbor terminal environments stated by repo. Must oracle-test constraints and image availability; preview format/revisions may change. https://github.com/icip-cas/LiteCoder.

Note RL task dataset + synthesizer | Teacher: LLM-assisted automatic synthesis; per-task creator metadata needed | Terms: Apache-2.0 dataset/code; image/package terms separate | Use: Train RL / scale new task synthesis | Paper reports >4,500environments; older releases1,376 must not be conflated. Main dataset ungated; seta-env-final separately gated. GitHub https://github.com/camel-ai/seta-env supplies Terminal-Bench/Harbor formats.

Note Existing terminal task-data reference, preserved. Public HF metadata/file inventory checked; no explicit license in card metadata. Files include container definitions and test_initial_state.py, but image/verifier integrity and origin/teacher terms are not runtime-validated. Related recipe: https://github.com/kanishkg/endless-terminals

Note Expert next-action RL/SFT data, NOT a verified release of all underlying terminal worlds. GLM-5.1 teacher; CC-BY-4.0. 31,111 per-turn samples from 630 seed tasks. Reward matches reference next actions via terminus_judge; split by original task/trajectory.

Note SFT multi-scaffold trajectories | Teacher: Teacher model unresolved in reviewed card; Claude Agent SDK used in task construction | Terms: MIT dataset; upstream seeds/teacher terms | Use: SFT long-horizon terminal | 11,255trajectories,10categories; Terminus2/OpenHands/ClaudeCode scaffolds. A ClaudeCode scaffold does not identify actual teacher. Approx27turns average; high rollout length.

Note SFT trajectories or supervised task records | Teacher: QuantTrio/GLM-4.6-AWQ with Terminus-2, explicitly named in card | Terms: apache-2.0 HF metadata; source/teacher terms still require review | Use: Consider selected train subsets after row-level success and provenance QC | Selected processed trajectories Current check establishes metadata and file presence, not task correctness or executable runtime. Mirrors/derived rows are not independent coverage.

Note RIGHTS HOLD. Historically used terminal SFT source, but current inspected metadata has no dataset license declaration. A teacher name in the title is not complete provenance. Obtain licensing/teacher/source confirmation and success evidence before new ingestion.