Third-party notices and attribution
The canonical library is an aggregate. The package source code license does not
replace source-dataset terms. Every record retains source_id, immutable
revision and record_locator; the evidence and conditions used for publication
are machine-readable in data/registry/permissions.json.
Publicly hosted payloads use the following source terms:
| Source | Terms and attribution |
|---|---|
areal_tau2_sft |
inclusionAI/AReaL-tau2-data, Apache-2.0. Credit inclusionAI and the SEA/AReaL authors, retain the licence notice and identify exact-prefix reconstruction and turn-target transformations. Respect underlying benchmark and teacher-output terms. |
litecoder_terminal_sft |
Lite-Coder/LiteCoder-Terminal-SFT, MIT. Credit the LiteCoder authors, retain the MIT licence notice and embedded third-party notices. Unknown teacher-output restrictions are not relabelled as verified commercial permissions. |
swegym_openhands |
SWE-Gym/OpenHands-SFT-Trajectories, MIT. Credit the SWE-Gym authors and retain the MIT notice. |
nvidia_open_swe_traces |
NVIDIA Open-SWE-Traces, CC-BY-4.0. Credit NVIDIA/Open-SWE-Traces, link CC-BY-4.0, identify canonical transformations, and retain the per-record repository license. |
terminaltraj |
m-a-p/TerminalTraj, linked publisher project under Apache-2.0. Credit the TerminalTraj authors and honor embedded MIT/Apache/BSD repository terms. |
official_tau3 |
Sierra Research tau2-bench, MIT. Copyright Sierra Research; retain the MIT notice. Training on benchmark tasks/rewards must be disclosed as evaluation overlap. |
mind2web |
OSU-NLP Mind2Web, CC-BY-4.0. Credit the Mind2Web authors, link CC-BY-4.0, identify transformations, and retain the research-purpose/harm disclaimer. |
toolbench |
OpenBMB ToolBench, Apache-2.0 with research-and-education intended use. Credit OpenBMB/ToolBench and consider third-party API content rights. |
toolbench_full |
OpenBMB ToolBench, transported from the pinned Apache-2.0 Parquet conversion. Retain Apache-2.0 attribution, the stated research/education intended use and third-party API-content caveat. |
delm_deepswe and delm_terminalbench4 |
DeLM 720 trajectories, MIT. Retain the MIT notice and benchmark-protection warnings. The embedded DeepSWE/Harbor canaries explicitly prohibit training-corpus inclusion, so these payloads are evaluation-only. |
swebench_verified_trajectories |
tarsur385/swebench-verified-trajectories, MIT. Retain the dataset notice; task repositories retain their own terms. |
swebench_multilingual_trajectories |
AlienKevin/SWE-bench-Multilingual-trajectories, MIT. Retain the dataset notice; task repositories retain their own terms. |
terminalbench_trajectories |
yoonholee/terminalbench-trajectories, Apache-2.0. Credit the dataset publisher and Terminal-Bench; consider the terms of model outputs represented in public leaderboard data. |
swesmith_trajectories |
SWE-bench/SWE-smith-trajectories, MIT. Credit SWE-smith and retain generated task repository terms. |
License references: MIT, Apache-2.0, and CC-BY-4.0.
The raw payloads for agentsuite_tau2, swegym_sampled,
browsecomp_rollouts, agentsuite_bfcl_v4,
swebench_multilingual_glm5, and
swebench_multilingual_minimax_m25 are metadata-only in the public aggregate
because the separate trajectory repositories do not establish sufficiently
clear transformed-redistribution terms. Their absence is not a claim that
access, research use, or every form of training is prohibited. The gated
alfworld_replayed_experts candidate is also metadata-only and was not
imported because the configured publishing identity has not accepted its
access terms.
This file is an engineering attribution aid, not legal advice.