Note SFT trajectories or supervised task records | Teacher: Not established from inspected metadata; ingestion requires teacher/source manifest | Terms: apache-2.0 HF metadata; source/teacher terms still require review | Use: Consider selected train subsets after row-level success and provenance QC | BFCL and tau2 clone-environment rows excluded by adapter Current check establishes metadata and file presence, not task correctness or executable runtime. Mirrors/derived rows are not independent coverag

Note SFT dataset; generation recipe | Teacher: APIGen synthetic pipeline; exact teacher mix not reverified in gated card | Terms: CC-BY-4.0 dataset; acknowledgement-gated files | Use: Train split SFT and schema/name/argument adaptation; decontaminate before reward conversion | Public card; unauthenticated README download returned401. Approval/citation acknowledgement required; full file integrity and all executable functions not tested.

Note RESTRICTED RESEARCH LANE. Multi-turn SFT, not a turnkey RL environment. CC-BY-NC-4.0; card also includes competitive-use warnings for GPT-generated content. Teachers: GPT-4o and DeepSeek-V3. Audit these terms and duplicated ancestry before use.

Note SFT dataset | Teacher: Tool-calling simulation/judge explicitly DeepSeek-V3.2 and GLM-4.6; audit other subsets separately | Terms: CC-BY-4.0; card additionally lists Apache-2.0/MIT ancestry | Use: Teacher-scoped tool/interactive subsets; direct train provenance | Underlying tool seeds include several other listed datasets; dedup and eval-domain filtering mandatory. Do not infer every subset has identical teacher provenance.

Note SFT dataset | Teacher: Qwen3-32B; Kimi-K2; GPT-OSS-120B listed separately | Terms: Apache-2.0 dataset card | Use: Select teacher-specific configurations and deduplicate against existing SFT | Current card/config inventory differs from historical OSS/SFT paths; pin exact revision. Large teacher traces need quality/length control.

Note SFT trajectories or supervised task records | Teacher: Not established from inspected metadata; ingestion requires teacher/source manifest | Terms: apache-2.0 HF metadata; source/teacher terms still require review | Use: Consider selected train subsets after row-level success and provenance QC | General_Agent and Search_Agent subsets Current check establishes metadata and file presence, not task correctness or executable runtime. Mirrors/derived rows are not independent coverage. Tiered-data fra

Note SFT candidate: prefer independently audited graph_syn rows. Apache-2.0 top-level card does not clear reconstructed upstream APIGen-MT NC or tau-bench ancestry. Teacher manifest, task-parent dedup and benchmark decontamination are required.

Note RL task data, not merely SFT traces. Apache-2.0 HF declaration. Pair with EnvFactory MCP implementations; audit gold/no-op/wrong-action rewards and reset behavior. Exclude benchmark-clone tasks and review generated-code/provider terms. https://github.com/LARK-AI-Lab/EnvFactory

Note SFT trajectories or supervised task records | Teacher: Not established from inspected metadata; ingestion requires teacher/source manifest | 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 examples Current check establishes metadata and file presence, not task correctness or executable runtime. Mirrors/derived rows are not independent coverage. Company policies limit rel

Note SFT/DPO/evaluation datasets and generation scripts | Teacher: Synthetic; generator accepts OpenAI-compatible endpoint/model; exact shipped teacher unestablished here | Terms: CC-BY-4.0 data; NVIDIA generation-code repo Apache-2.0 | Use: Use training/preference splits; hold benchmark split out | No-call correctness is not merely matching a function. Public generator can make new train domains. Audit shipped teacher before open-only training.

Note train/test dataset; executable API benchmark | Teacher: Multi-agent synthesis; human/LLM origins need sample-level audit | Terms: MIT dataset card; implementation repo alsoMIT | Use: Training portion SFT; new disjoint tasks can support execution reward after adapter | HF viewer currently cast-fails because files have differing schemas; read train/test files separately. Synthetic health/finance examples need safety review.

Note RL task dataset with DB/KB and executable Gym environment | Teacher: Synthetic banking worlds/tasks; exact task-generation teacher not stated | Terms: Apache-2.0 dataset card | Use: 250train/50validation tasks; can construct disjoint customer/tasks safely | Mostly English; not evidence of 22-language coverage. Judge/simulator are configured services, not dataset weights.

Note SFT candidate. Apache-2.0 card; exact shipped teacher provenance needs review. We historically used only the no-call slice. Tool-call examples require fresh syntax/argument/result QC; public demonstrations do not guarantee an executable tool environment.

Note Keep benchmark cases held out. training and evaluation dataset | Teacher: GPT-4 synthesis/quality checks; source dialogues include MultiWOZ and SGD | Terms: Code Apache-2.0; HF data card lacks license declaration in inspected metadata | Use: MTU-Instruct only for SFT after data-license approval; MTU-Eval held out | MTU means multi-granularity, not multilingual. Code license cannot automatically license source dialogue data.

Note RL prompt/task dataset | Teacher: Synthetic; exact generation teacher not established | Terms: CC-BY-4.0 dataset card | Use: Existing train source or new world/domain variants | Task rows require tool/environment implementation, not standalone SFT. Separate final-state correctness from only matching action traces.

Note RL task dataset | Teacher: Synthetic; exact teacher unspecified | Terms: CC-BY-4.0 dataset card | Use: Train prompts plus verified scheduling reward; can extend horizon and conflict density | Long prompt/output parsing can dominate CPU runtime. Gold parser must handle reasoning tags/JSON faithfully.

Note RL task dataset | Teacher: DeepSeek-V3.2, GLM-4.6, GPT-OSS-120B, Kimi-K2-Instruct listed | Terms: CC-BY-4.0 card; mentions Apache-2.0/MIT ancestry | Use: Training with controlled distractors and disjoint schemas | Comparator configuration defines handling of parallel/extraneous calls; pin evaluator implementation.

Note RL task dataset | Teacher: DeepSeek-R1-0528/V3.2, Qwen3Thinking/32B listed; judges GPT-OSS120B/Qwen3Instruct | Terms: CC-BY-4.0 dataset card | Use: Train with multi-turn state/history and wrong-argument negatives | Static action rows are not full live user simulator episodes. Correct histories and action boundaries must remain intact.

Note EVAL ONLY for original benchmark cases. Cross-domain interactive services; MIT dataset declaration. Preserve original task IDs for held-out assessment. New independently authored worlds require simulator, state-reset and oracle checks.

Note RIGHTS HOLD / EVAL ONLY. Sequential user-memory benchmark; original sequences must remain held out. HF task-data license is absent in inspected files despite MIT code. Do not infer dataset permissions from the implementation license.

Note EVAL ONLY for original cases. Apache-2.0 dataset; multi-step constrained and long-context tool tasks. Keep benchmark questions held out; runtime/API stability and actual model context support must be verified.

Note EVAL ONLY for original queries. Multi-hop tool benchmark with executable tool code. README declares data CC-BY-4.0 and code Apache-2.0 despite missing HF license tag. Split new training tasks by parent MorehopQA question; sandbox tool code.

Note Keep benchmark cases held out. train/validation/test planning dataset; DB/evaluator | Teacher: Task generation provenance not established; original output parser uses GPT-4 | Terms: Dataset CC-BY-4.0; codeMIT | Use: Training split and released finetuning_data after teacher review; never train held-out test | Not native Gym; write scoped wrapper. Plan parser/judge choice affects reward; canonical JSON avoids paid parser as separate change.

Note synthetic train/dev/test dataset | Teacher: Adaptive Data by Adaption synthetic pipeline; teacher model not specified | Terms: Apache-2.0 HF card/README claim; no separate GitHub LICENSE file found | Use: Small Hindi/Punjabi train-split candidate after native-speaker/schema audit | 10551records/12languages; high-stakes health/finance examples need safety review. README quality claims not independently validated.

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 | Tool-calling and interactive-agent subsets Current check establishes metadata and file presence, not task correctness or executable runtime. Mirrors/derived rows are not independent coverage.

Note SFT trajectories or supervised task records | Teacher: Not established from inspected metadata; ingestion requires teacher/source manifest | Terms: apache-2.0 HF metadata; source/teacher terms still require review | Use: Consider selected train subsets after row-level success and provenance QC | Only func_calling configuration; Glaive-derived configurations excluded Current check establishes metadata and file presence, not task correctness or executable runtime. Mirrors/derived rows are not inde

Note SFT trajectories or supervised task records | Teacher: Not established from inspected metadata; ingestion requires teacher/source manifest | Terms: apache-2.0 HF metadata; source/teacher terms still require review | Use: Consider selected train subsets after row-level success and provenance QC | LoopTool_grpo_training_data.json adapted to one-turn SFT records Current check establishes metadata and file presence, not task correctness or executable runtime. Mirrors/derived rows are not independent

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 | Irrelevant-tool examples Current check establishes metadata and file presence, not task correctness or executable runtime. Mirrors/derived rows are not independent coverage.

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 | Clarification examples Current check establishes metadata and file presence, not task correctness or executable runtime. Mirrors/derived rows are not independent coverage.

Note SFT trajectories or supervised task records | Teacher: Not established from inspected metadata; ingestion requires teacher/source manifest | Terms: apache-2.0 HF metadata; source/teacher terms still require review | Use: Consider selected train subsets after row-level success and provenance QC | Only examples asking for missing argument values Current check establishes metadata and file presence, not task correctness or executable runtime. Mirrors/derived rows are not independent coverage.