data/DATASETS.md

# DADOES Dataset Survey

DADOES detects mood from English daily reports written by agents in a
simulated world. Public emotion corpora are useful for a foundation model, but
they do not fully cover the DADOES taxonomy. The critical gap is direct
supervision for `lonely`, `bored`, and `tired`.

Last checked: 2026-07-01.

## Coverage Summary

| Mood area | Public coverage | Recommendation |
|---|---|---|
| Broad positive and negative moods | Good coverage from GoEmotions and related emotion corpora. | Keep GoEmotions as the public baseline. |
| `lonely` | Usable public datasets exist, mostly Reddit-derived. | Add a separate loneliness adapter/evaluation path after license review. |
| `bored` | No clean English text emotion dataset found with reliable boredom labels. | Build DADOES-owned labels from agent reports. |
| `tired` | Public fatigue datasets are mostly medical, driving, physiological, or non-emotion tasks. | Build DADOES-owned labels from agent reports. |
| Agent daily-report style | Public corpora are mostly Reddit, Twitter, or dialogue. | Collect and label in-domain reports. |

## Candidate Datasets

| Dataset | License / status | Useful labels | DADOES decision |
|---|---|---|---|
| [`google-research-datasets/go_emotions`](https://huggingface.co/datasets/google-research-datasets/go_emotions) | Apache-2.0 | Broad fine-grained emotions plus neutral. No direct `lonely`, `bored`, or `tired`. | Already used as the baseline training and evaluation source. |
| [`facebook/empathetic_dialogues`](https://huggingface.co/datasets/facebook/empathetic_dialogues) | CC-BY-NC-4.0 | Dialogue situations include `lonely` and several useful adjacent moods. | Parser is wired for local exported CSV, but trainer only reads it when `DADOES_INCLUDE_NON_COMMERCIAL=1`. Do not ship public weights trained on it as AGPL-only artifacts. |
| [`FIG-Loneliness/FIG-Loneliness`](https://huggingface.co/datasets/FIG-Loneliness/FIG-Loneliness) | CC-BY-NC-4.0 | Binary loneliness plus fine-grained loneliness context fields. | Parser is wired for local exported JSONL, but trainer only reads it when `DADOES_INCLUDE_NON_COMMERCIAL=1`. |
| [`yael-katsman/Loneliness-Causes-and-Intensity`](https://huggingface.co/datasets/yael-katsman/Loneliness-Causes-and-Intensity) | CC-BY-4.0 | Loneliness causes and intensity from 1 to 5. Small dataset. | Public optional source. The trainer maps it to partial `lonely` supervision. |
| [`smit18/text_emotion`](https://huggingface.co/datasets/smit18/text_emotion) | Apache-2.0, sparse dataset card | Twitter-style labels: `happiness`, `love`, `sadness`, `worry`, `hate`. | Public optional source. It is auxiliary only and does not cover `lonely`, `bored`, or `tired` as labels. |
| [`PrajwalNayaka/Text-Emotion`](https://huggingface.co/datasets/PrajwalNayaka/Text-Emotion) | Apache-2.0, sparse dataset card | `neutral`, `happiness`, `love`, `empty`, `hate`, `anger`, `enthusiasm`, `relief`, `fun`, `sadness`, `surprise`. | Public optional source with sparse provenance metadata. `empty` maps to `sad`, not `lonely`, to avoid false direct supervision. |
| [`Um1neko/text_emotion`](https://huggingface.co/datasets/Um1neko/text_emotion) | Apache-2.0 metadata | Plutchik-like labels with intensity: anger, anticipation, disgust, fear, joy, sadness, surprise, trust, neutral. | Public optional source for broad mood coverage. No direct missing dimensions. |
| DailyDialog, EmotionLines, MELD, EmoWOZ | Varies by source | Conversation emotions, usually basic emotions or task-dialogue sentiment. | Useful for conversation robustness, not for `bored` or `tired` supervision. |
| DADOES agent reports | Project-owned if collected from our simulation. | All DADOES labels, especially `lonely`, `bored`, and `tired`. | Required for production-quality coverage. |

## Required DADOES Domain Set

Public data should not be treated as enough for the missing dimensions. The
first production data set should be project-owned and should include:

- At least 100 held-out reviewed examples each for `lonely`, `bored`, and
  `tired`.
- At least 50 held-out reviewed examples for every other mood.
- Multi-label records, because a report can be both `lonely` and `tired`, or
  `bored` and `frustrated`.
- Negative examples for each missing dimension. For example, a report that says
  "I was alone but focused and calm" should not automatically be `lonely`.
- Source metadata outside the model input: world day, agent id, location, and
  major events. This lets us audit bias without leaking metadata into the
  text-only classifier.

Recommended JSONL schema:

```json
{"text":"I patrolled the empty road for hours and nobody answered my calls.","labels":["lonely","tired"],"source":"agent_report","reviewed":true}
```

## Label Rules For Missing Dimensions

Use these rules when labeling agent reports:

- `lonely`: social isolation, missing companionship, rejection, abandonment, or
  inability to reach others. Being physically alone is not enough.
- `bored`: under-stimulation, repetitive routine, lack of meaningful activity,
  waiting without engagement, or explicit boredom. Calm is not boredom.
- `tired`: physical or mental fatigue, exhaustion, sleepiness, depletion, or
  inability to continue from low energy. Sadness or frustration is not enough.

## Training Plan

1. Keep the current GoEmotions checkpoint as the public baseline.
2. Use `cargo run --release --features dataset-download --bin
   prepare_external_datasets` to prepare public raw-file sources under
   `data/raw/external`.
3. Keep non-commercial sources behind `DADOES_INCLUDE_NON_COMMERCIAL=1` and in a
   separate model card. Provide their local CSV/JSONL exports explicitly.
4. Build `domain_eval.jsonl` before using synthetic or weakly labeled domain
   data for training.
5. Use synthetic and LLM-labeled reports only for training expansion, not for
   final benchmark claims unless they are human-reviewed.