An executable Environment for tool-using agents, rebuilt from recorded traces. Release: 100.0% (130 of 130) of its Tasks replay as recorded; a release needs 90%. fetch checks the content hash; --revision fetches an earlier publish. harbor/ holds 123 Tasks in the Harbor task format; a judge item is scored at test time by the model KULLBACKJUDGEMODEL names, with HARNESSALLOWMODELREQUESTS=1 and the model's keys; without an answer the trial is unscored, not 0. 130 Tasks over 13 tools, each with the world it starts from and, where one was derived, its Verifier in code; tasks.jsonl lists them. One Task's instruction opens, for example: "wants to book a flight from San Francisco to New York for 3…
Organization
Leibler
leibler
RL Envs and Post training models
Compare the published evidence
This profile connects the 2 model and dataset records shown below to their upstream cards and sources. The library counts describe SAVRN's coverage, rather than this publisher's entire output.
Reported licenses: mit. Open a record to review its specific license and conditions.
Compare model specifications, source dates, and hardware estimates on the individual records. Publisher identity verification is separate from a model's quality, deployment readiness, or suitability for your work.
Datasets
An executable Environment for tool-using agents, rebuilt from recorded traces. Release: 100.0% (223 of 223) of its Tasks replay as recorded; a release needs 90%. fetch checks the content hash; --revision fetches an earlier publish. harbor/ holds 186 Tasks in the Harbor task format; a judge item is scored at test time by the model KULLBACKJUDGEMODEL names, with HARNESSALLOWMODELREQUESTS=1 and the model's keys; without an answer the trial is unscored, not 0. 223 Tasks over 16 tools, each with the world it starts from and, where one was derived, its Verifier in code; tasks.jsonl lists them. One Task's instruction opens, for example: "wants to return the laptop order for a full refund because…