ConTSG-Bench Specification (Single Source of Truth)

March 4, 2026 ยท View on GitHub

This document is the canonical reference for benchmark scope. Public-facing docs should reference this file instead of duplicating counts.

Last updated: 2026-03-03

Scope Summary

  • Benchmark datasets: 10
  • Generation models in benchmark suite: 11
  • Leaderboard metrics: 15
    • Fidelity: 7
    • Adherence: 4
    • Utility: 4

Benchmark Dataset IDs (10)

  1. synth-m
  2. synth-u
  3. ettm1
  4. weather_concept
  5. weather_morphology
  6. telecomts_segment
  7. istanbul_traffic
  8. airquality_beijing
  9. ptbxl_concept
  10. ptbxl_morphology

Generation Model IDs (11)

Text-conditioned

  • verbalts
  • t2s
  • bridge
  • diffusets
  • text2motion
  • retrieval

Attribute-conditioned

  • timeweaver
  • wavestitch
  • tedit

Label-conditioned

  • timevqvae
  • ttscgan

Leaderboard Metric IDs (15)

Fidelity (7)

  • acd
  • sd
  • kd
  • mdd
  • fid
  • prdc_f1.precision
  • prdc_f1.recall

Adherence (4)

  • jftsd
  • joint_prdc_f1.precision
  • joint_prdc_f1.recall
  • cttp

Utility (4)

  • dtw
  • crps
  • ed
  • wape

Ranking Policy Notes

  • Overall ranking uses Fidelity + Adherence groups.
  • Utility metrics are reported but excluded from overall ranking.