Laya Task Unified v2
Fine-tuned from convaiinnovations/laya to perform both stages of a task-management pipeline: classify a task as home, work, or personal, then apply the corresponding area-specific tags.
This checkpoint was trained on the unified configuration of chand1012/laya-task-routing-synthetic-v2.
Evaluation
On the held-out synthetic v2 test split (363 records, 2,001 decisions):
| Metric | Result |
|---|---|
| Decision accuracy | 94.80% |
| Record exact match | 76.31% |
| Area accuracy | 95.58% |
| Area macro-F1 | 95.53% |
| Tag micro-F1 | 86.81% |
| Tag exact-set accuracy | 57.14% |
See evaluation_report.json for calibration, latency, per-area, and per-tag results. These are synthetic-set results and should not be interpreted as production performance.
Usage
import laya
agent = laya.load("chand1012/laya-task-unified-v2")
result = agent.predict(state, questions)
The caller supplies the same typed Laya questions used during training. The taxonomy and record schema are documented in the dataset repository.
Limitations
The training and evaluation data are synthetic, English-only, and constrained to a three-area taxonomy with ten tags per area. Validate on real, independently labeled tasks before production use. Performance can degrade on ambiguous, multilingual, adversarial, or out-of-taxonomy inputs.
License
Apache License 2.0. This is a derivative of the Apache-2.0-licensed Laya base model.