| language: | |
| - en | |
| license: mit | |
| task_categories: | |
| - reinforcement-learning | |
| configs: | |
| - config_name: default | |
| data_files: | |
| - split: train | |
| path: tasks.jsonl | |
| Original QE tasks cover subtitles, timecodes, catalogues, assets and exact rational ledger aggregation. Reward is one when all original output-artifact checks pass, otherwise zero. | |
| Build `components/qefiles/Dockerfile` and run the image on port 8000. Reset with `task_id` set to `qe_skill-media-d2` or `qe_skill-math-d4`; an explicit seed reproduces an instance. Actions are `exec`, `write_file` and `evaluate`. Use one active session per sandbox. | |
| Code and synthetic task data are by Dragan Spiridonov, licensed under MIT. Source and SDK revisions are recorded in `components/qefiles/SOURCE_MANIFEST.json`; original notices are retained. Run `python3 -B validate_dataset.py` to verify the inventory. | |