README.md

---
license: apache-2.0
task_categories:
  - text-classification
language:
  - en
pretty_name: Agent Task Classifier Dataset (laya loop-taxonomy v2)
size_categories:
  - 1K<n<10K
tags:
  - agent
  - routing
  - task-classification
  - decision-model
  - laya
---

# agent-task-classifier-dataset

5,154 labeled English training examples (plus 54 held-out) for classifiers that map a user
request (to a personal-assistant agent) to the **agent-loop shape** it demands — the skeleton of
the high-level process a task of that kind must follow — rather than its topic.

Companion model: [kumarabd1996/agent-task-classifier](https://huggingface.co/kumarabd1996/agent-task-classifier)
(a fine-tuned [laya](https://huggingface.co/convaiinnovations/laya) checkpoint; 47/54 on the
held-out set here, vs 13/54 for the untuned base).

**v2** relabels the v1 corpus onto a nine-label set and adds ~1,500 personal-records and finance
scenarios (goals, routines, budgets, trackers and the lookups around them). The v1 files are in this
repo's git history.

## The nine labels

| label | the loop it demands |
|---|---|
| `conversation` | questions, quick asks, lookups, a reported measurement, a single reminder, chatter — the default, safe fallback |
| `companion` | inner-life support with no outcome in the world (reflection, therapy) |
| `learn` | build the user's own ability over sessions (practice, tutoring) |
| `diagnose` | find the cause of a specific malfunction or anomaly, including in the user's own data |
| `research` | assemble understanding from investigation (multi-source, comparison, synthesis) |
| `produce` | make an artifact judged against a standard (document, spreadsheet, script, letter) |
| `transact` | a commitment or possession in the world: book, buy, pay, transfer, or complete a procedure |
| `stabilize` | urgent or deteriorating situation — act before full understanding |
| `plan` | set up or change the person's own standing structure: a goal, routine, budget/limit, or tracker |

v1 → v2: `make`→`produce`; `react`→`stabilize`; `arrange`/`acquire`/`execute`→`transact`;
`tend`→`conversation`; `plan` is new. Two requests share a label iff they demand the same loop
skeleton. Topic is irrelevant.

## Format

`train.jsonl` — one JSON object per line, laya's native training format:

```json
{"state": "<the request text>", "questions": {"loop": {"type": "choice", "instructions": "...", "criteria": {...}}}, "expected": {"loop": "<label>"}, "src": "<source>", "id": "<id>"}
```

Continuations carry the request being continued, in the exact format used at inference:

```
Earlier request: <prior request>
New request: <follow-up>
```

`eval.jsonl` — 54 held-out items: the 24 taxonomy boundary drills (relabelled) plus 30 hand-written
personal-records and finance cases (`eval_life_finance.json`, phrased differently from any template).
Never trained on.

## Composition

| source (`src`) | what | license |
|---|---|---|
| `multiwoz`, `multiwoz-cont` | MultiWOZ 2.1 booking utterances + continuations (`transact`) | permissive (research; Budzianowski et al. 2018) |
| `massive` | MASSIVE en-US takeaway orders, transport bookings, sampled commands | CC-BY-4.0 (Amazon Science) |
| `oasst`, `oasst-cont` | OASST2 English prompts + continuations, model-labeled with the taxonomy spec, spot-audited | Apache-2.0 (OpenAssistant) |
| `synth`, `synth-cont`, `hard` | template × slot generation across the personal-assistant domains; labels by construction; hand-written boundary cases | Apache-2.0 (this release) |
| `lf`, `lf-cont`, `lf-hard` | personal-records and finance scenarios incl. short follow-ups ("All", "yes"); labels by construction | Apache-2.0 (this release) |

Label distribution (after per-class caps): conversation 1,300 · transact 1,200 · plan 633 · produce 600 · diagnose 551 · research 333 · learn 226 · companion 187 · stabilize 124.

## Provenance and licensing

- **MASSIVE**: Amazon Science, CC-BY-4.0 — attribution required and given.
- **OASST2**: OpenAssistant Conversations Dataset, Apache-2.0.
- **MultiWOZ 2.1**: Budzianowski et al., 2018; distributed for research use.
- **Synthetic items**: generated for this release; scripts included (`generate_synthetic.py`,
  `generate_life_finance.py`, `merge_v2.py`, `taxonomy_labels.py`) so the construction is auditable.
  `generate_synthetic.py` still emits the v1 label names; `merge_v2.py` maps them to v2 (its inputs, the
  v1 corpus files, are in this repo's git history, not in the current tree).

## Limitations

- A large share is synthetic; template patterns bias the text distribution. The `plan` examples in
  particular are template-built, so implicit phrasings ("I keep meaning to exercise more…") are
  underrepresented.
- English only; nine fixed labels; out-of-taxonomy requests resolve to `conversation`.
- OASST2 labels are teacher-model labels (spot-audited), not fully human-verified. The MultiWOZ
  "find a hotel" requests are labeled `transact`, a judgment call.
- 54 held-out items is a small evaluation set.
- **Privacy**: public datasets and synthetic text only. The maintainer's own data is excluded.