open-athena/nemotron-gym-structured-outputs-v4
laion/nemotron-gym-structured-outputs-v4 Harbor task-binary dataset (53,870 tasks) converted from nvidia/Nemotron-RL-Instruction-Following-Structured-Outputs-v2 (part of nvidia/Nemotron-Post-Training-v3). Columns path (str) + task_binary (gzip tar). Converted with the OpenThoughts-Agent data.nemotron_gym framework. Grading: JSON/YAML/TOML schema validation; XML/CSV structural. What changed vs the prior version This version fixes the answer-delivery contract for… See the full description on the dataset page: https://huggingface.co/datasets/open-athena/nemotron-gym-structured-outputs-v4.
laion/nemotron-gym-structured-outputs-v4
Harbor task-binary dataset (53,870 tasks) converted from nvidia/Nemotron-RL-Instruction-Following-Structured-Outputs-v2 (part of nvidia/Nemotron-Post-Training-v3).
Columns path (str) + task_binary (gzip tar). Converted with the OpenThoughts-Agent data.nemotron_gym framework.
Grading: JSON/YAML/TOML schema validation; XML/CSV structural.
What changed vs the prior version
This version fixes the answer-delivery contract for terminal agents. The prior version told the agent what to produce but not how to submit it; a 1-turn terminus-2 agent emitted its answer as a chat reply instead of writing the graded file, so most trials scored 0 with "answer file missing". The instruction now explicitly instructs writing to the grader's file path via a shell heredoc (and verifying it). Grading logic is otherwise unchanged.