DADOES Data

DADOES starts from GoEmotions-style multi-label mood detection, but the target domain is an English daily report written by an agent in a simulated world.

Training records should use JSON Lines:

{"text":"I repaired the shelter and felt satisfied before sunset.","labels":["satisfied"]}
{"text":"The wolves followed me back, and I felt anxious and tired.","labels":["anxious","tired"]}

The first production dataset should combine:

GoEmotions-derived public examples mapped into the DADOES label set.
License-reviewed public loneliness examples.
Agent daily reports sampled from the world.
Human or LLM-reviewed labels for those reports.

The public dataset survey is tracked in DATASETS.md. The practical decision from that survey is that lonely can be improved with external English text data, while bored and tired require project-owned agent-report labels.

Optional external sources are read from data/raw/external:

data/raw/external/
  dadoes-domain/train.jsonl
  dadoes-domain/dev.jsonl
  dadoes-domain/test.jsonl
  loneliness_causes/train_data.csv
  loneliness_causes/validation_data.csv
  loneliness_causes/test_data.csv
  text_emotion/text_emotion.csv
  prajwal_text_emotion/final_dataset.csv
  um1neko/train.json
  um1neko/dev.json
  um1neko/test.json
  empathetic_dialogues/train.csv
  empathetic_dialogues/validation.csv
  empathetic_dialogues/test.csv
  fig_loneliness/train.jsonl
  fig_loneliness/dev.jsonl
  fig_loneliness/test.jsonl

Run cargo run --release --features dataset-download --bin prepare_external_datasets from the repository root to download the public raw-file sources.

The Rust trainer can also read locally exported non-commercial sources:

empathetic_dialogues/*.csv must contain context, prompt, and utterance columns.
fig_loneliness/*.jsonl must contain text and lonely fields.

DADOES does not depend on Python or Hugging Face datasets to produce those files. The trainer requires DADOES_INCLUDE_NON_COMMERCIAL=1 before it will read them.

The Rust baseline in src/lib.rs uses a small in-crate seed set only so the project has a runnable machine-learning path from the first commit.