Note RL task dataset + generator | Teacher: Synthetic mutations/issue descriptions; generator model varies | Terms: MIT dataset/toolkit; underlying repository licenses remain | Use: Train RL; create fresh repository-disjoint tasks | 59,136 train rows in current API versus older README 52K. Pin revision. Docker images and oracle replay required; never expose gold patch/tests to policy.

Note SFT trajectories | Teacher: Claude 3.7 Sonnet (proprietary) | Terms: MIT dataset; upstream code and teacher terms separately | Use: SFT selected successful tool-format trajectories | tool/xml/ticks are alternate encodings of related trajectories, not independent task diversity. Current API ~76K records across formats; earlier card prose references5,017. Dedup by task/trajectory.

Note RL task dataset + executable environments | Teacher: Real GitHub PRs/issues; LLM annotations/build assistance (builder supports GPT-4.1-mini) | Terms: CC-BY-4.0 data; MIT builder; per-row repository SPDX varies | Use: Train on explicit train; multilingual programming languages | 32,079 train tasks /20 programming languages. Explicit warning: original SWE-bench repos removed, other benchmark overlap NOT fully removed. Read overlap list before training; downloaded image integrity untested.

Note SFT trajectories | Teacher: Qwen3-Coder-480B-A35B-Instruct; OpenHands0.54 | Terms: CC-BY-4.0; underlying repositories/teacher terms | Use: SFT / on-policy-independent warm start | Convert stringified trajectory/tool arguments consistently; dedup against Nemotron SWE and other task derivatives; a trace is not an executable environment.

Note RL task dataset + runnable gym | Teacher: SWE-GEN synthetic task and verifier construction from commits | Terms: Apache-2.0 dataset/gym; repository licenses separate | Use: Train RL / test generation | 8,101 current train rows. Repo https://github.com/R2E-Gym/R2E-Gym. Validate generated tests for specificity, nondeterminism and leaks; original13-repo concentration limits breadth.

Note RIGHTS HOLD. Claude-generated SWE SFT trajectories. Dataset license is not explicitly declared in inspected metadata; Apache-licensed R2E-Gym code does NOT settle data rights. Audit source/teacher terms and success before training.

Note RL task dataset + executable gym | Teacher: Human GitHub tasks with curated execution environments | Terms: MIT dataset; Apache-2.0 gym code; repository licenses separate | Use: Train RL / verifier learning | 2,438 train tasks,11Python repos. https://github.com/SWE-Gym/SWE-Gym. Repo-disjoint validation essential; test failures can mean image breakage, not model failure.

Note SFT data (agentic + agentless) | Teacher: Qwen3-Coder480B OpenHands; DeepSeek-R1-0528 agentless | Terms: CC-BY-4.0; underlying Apache/MIT/BSD notices | Use: SFT agentic split plus separately budgeted agentless repair | ~46K OpenHands trajectories cited in card; agentless is not multi-turn agent data. Source tasks include SWE-Gym/R2E and SWE-Smith/rebench; decontam lineage, not only text.

Note SFT multi-scaffold trajectories | Teacher: MiniMax/Qwen/DeepSeek families; current v1.2 Qwen3.8-27B | Terms: CC-BY-4.0; per-row MIT/Apache/BSD code provenance | Use: SFT across OpenHands/SWE-agent/mini-swe-agent | Pin revision/config. v1.0 reduced207K→151K after git-hacking cleanup; newest traces differ from old mixture. Do not equate scaffold named Claude Code with Claude teacher.

Note SFT candidate with GLM-4.6 teacher; ODC-BY dataset declaration and research/responsible-use guidance. Authors explicitly did NOT verify main experiment trajectories. Distinguish target patch from produced patch and execute tests before using traces.

Note SFT trajectories | Teacher: LLM-generated multi-turn interactions; model provenance needs paper-level row review | Terms: Apache-2.0 data/code; mixed upstream datasets and teacher terms | Use: SFT code-action control, NOT standalone full RL task pool | 7,139 codeact rows versus71,246 general rows. Budget codeact separately; avoid counting general conversations as agent episodes. Executable-code actions require sandbox/tool adapter.

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: DeepSeek V3.2 | Terms: README says CC-BY-4.0; HF license metadata absent; resolve LICENSE/upstream code | Use: Research SFT after license/provenance audit | 71K trajectories /3.5B tokens stated; highly overlapping GitHub tasks possible. Card claim is weaker than complete license-file audit; no runtime replay performed.

Note SFT trajectories or supervised task records | Teacher: Not established from inspected metadata; ingestion requires teacher/source manifest | Terms: cc-by-4.0 HF metadata; source/teacher terms still require review | Use: Consider selected train subsets after row-level success and provenance QC | question_tool, agent_skills, agent_skills_question_tool and general subsets Current check establishes metadata and file presence, not task correctness or executable runtime. Mirrors/derived rows are not i

Note SFT trajectories or supervised task records | Teacher: Not established from inspected metadata; ingestion requires teacher/source manifest | Terms: mit 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.