CoolFace
Datasetpublic

chand1012/laya-task-routing-synthetic-v2

Laya Task Routing Synthetic v2 Synthetic English task data for experiments that classify a task into an area (home, work, or personal) and then apply area-specific multi-label tags. This dataset was generated by a self-hosted Qwen 3.8 27B model and subsequently reviewed by the same model as part of the synthetic-data validation pipeline. Configurations Config Records Decisions Purpose area 1,800 1,800 Area classification only tags 1,800 18,000… See the full description on the dataset page: https://huggingface.co/datasets/chand1012/laya-task-routing-synthetic-v2.

sourceHugging Faceapache-2.0updated 2h agoView on Hugging Face
0likes
Dataset Card

Laya Task Routing Synthetic v2

Synthetic English task data for experiments that classify a task into an area (home, work, or personal) and then apply area-specific multi-label tags.

This dataset was generated by a self-hosted Qwen 3.8 27B model and subsequently reviewed by the same model as part of the synthetic-data validation pipeline.

Configurations

ConfigRecordsDecisionsPurpose
area1,8001,800Area classification only
tags1,80018,000Area-conditioned tagging only
unified3,60019,800Both decision roles in one training corpus

The three configurations contain train, validation (called calibration during training), and test splits. Semantic groups and near-duplicate clusters are kept within one split. The generation audit found no cross-split near-duplicate pairs at its 0.65 threshold and no tag-cardinality violations.

Generation

The corpus was generated and model-reviewed with a self-hosted Qwen 3.8 27B model through an OpenAI-compatible local endpoint. Each record includes provenance, difficulty, scenario family, rationale, and validation metadata. Generation used controlled scenario, phrasing, difficulty, and boundary-case variation, followed by structural validation and model review.

The complete taxonomy, class counts, tag-positive counts, generation metadata, and audit results are in manifest.json.

Loading

python
from datasets import load_dataset

area = load_dataset("chand1012/laya-task-routing-synthetic-v2", "area")
tags = load_dataset("chand1012/laya-task-routing-synthetic-v2", "tags")
unified = load_dataset("chand1012/laya-task-routing-synthetic-v2", "unified")

Intended use

This dataset is intended for research and prototyping of task routing and tagging models, including comparison of a unified model against two specialist models. It can also be used for regression tests and calibration experiments.

Limitations

  • —All examples are synthetic and may reflect generator-model biases, repeated patterns, or unrealistic phrasing.
  • —Synthetic test results do not estimate production performance on real user tasks.
  • —The taxonomy is deliberately narrow and may not fit every task-management system.
  • —Human review and evaluation on independently collected, consented real-world data are recommended before deployment.
  • —Rationales are generator-produced metadata and should not be treated as independently verified explanations.

Evaluation artifacts

laya-v2-evaluation-comparison.json contains the unified-versus-specialist comparison. Evaluation was performed on this synthetic test split, so the results should be treated as an internal benchmark rather than a real-world generalization claim.

License

Apache License 2.0.