Datasets:
| # TRACE data generation config | |
| # Controls per-area example counts, split ratios, and the global seed. | |
| seed: 42 | |
| output_dir: data/splits | |
| # Per-area generation counts and enabled flags. | |
| # Total target: ~3,500 examples (train + valid + curation-pool sums). | |
| areas: | |
| dtt: | |
| enabled: true | |
| n: 800 | |
| description: "DTT teaching programs — array-based discrete-response skills." | |
| net: | |
| enabled: true | |
| n: 500 | |
| description: "NET teaching programs — motivation-based naturalistic teaching." | |
| task_analysis: | |
| enabled: true | |
| n: 500 | |
| description: "Task Analysis / Chaining — multi-step routines (AFLS)." | |
| fct: | |
| enabled: false | |
| n: 300 | |
| description: "FCT — replacement-behavior programs paired with deceleration targets." | |
| bst: | |
| enabled: false | |
| n: 200 | |
| description: "BST — staff-facing training programs." | |
| prt: | |
| enabled: false | |
| n: 300 | |
| description: "PRT — pivotal-behavior programs." | |
| session_interpretation: | |
| enabled: true | |
| n: 1200 | |
| description: "Task 2 — behavioral session interpretation (paper's innovation task)." | |
| # Split ratios applied after all enabled areas generate | |
| # test + sanity are NOT generated here — they're curated by the user from curation_pool | |
| splits: | |
| train: 0.85 | |
| valid: 0.05 | |
| curation_pool: 0.10 # user reviews this pool and curates test.jsonl + sanity.jsonl | |