Two open datasets dramatically improve empathetic, norm-aware responses without fine-tuning: (1) EmpatheticDialogues (facebook/empathetic_dialogues on HuggingFace) -- 25K conversations grounded in 32 emotional situations, CC license, 25MB download. wget https://dl.fbaipublicfiles.com/parlai/empatheticdialogues/empatheticdialogues.tar.gz. Gives the model emotional grounding for parent Q&A. (2) Social Chemistry 101 (github.com/mbforbes/social-chemistry-101) -- 292K rules of thumb about social and moral norms, CC BY-SA 4.0. storage.googleapis.com/ai2-mosaic-public/projects/social-chemistry/data/social-chem-101.zip. Provides the unwritten social rules ASD families need made explicit. Work: (1) Download both datasets to a new data/external/ directory (gitignored), (2) write an ingestion script scripts/ingest_empathetic_dialogues.py and scripts/ingest_social_chemistry.py that convert each to the corpus JSONL chunk format, (3) add these as optional corpus sources configurable via --sources flag on AskASailorAgent. Tests: verify chunk format, verify no PII, verify chunk count.
Two open datasets dramatically improve empathetic, norm-aware responses without fine-tuning: (1) EmpatheticDialogues (facebook/empathetic_dialogues on HuggingFace) -- 25K conversations grounded in 32 emotional situations, CC license, 25MB download. wget https://dl.fbaipublicfiles.com/parlai/empatheticdialogues/empatheticdialogues.tar.gz. Gives the model emotional grounding for parent Q&A. (2) Social Chemistry 101 (github.com/mbforbes/social-chemistry-101) -- 292K rules of thumb about social and moral norms, CC BY-SA 4.0. storage.googleapis.com/ai2-mosaic-public/projects/social-chemistry/data/social-chem-101.zip. Provides the unwritten social rules ASD families need made explicit. Work: (1) Download both datasets to a new data/external/ directory (gitignored), (2) write an ingestion script scripts/ingest_empathetic_dialogues.py and scripts/ingest_social_chemistry.py that convert each to the corpus JSONL chunk format, (3) add these as optional corpus sources configurable via --sources flag on AskASailorAgent. Tests: verify chunk format, verify no PII, verify chunk count.