Algorithms
Summary of available tau algorithms (Run 3)
Legend:
- Fully supported for Run 3
- Pending calibration
- Development ongoing
- Superseded
| Status | Task | Description | Documentation |
|---|---|---|---|
|
DeepTau v2.5 | CNN architecture with domain adaptation (taus only) | TAU-24-001 |
|
UParT | Particle Transformer (unified jet tagging) | DP2024_066 |
|
ParticleNet | GNN (unified jet tagging) | AN-22-094 |
|
Tau Transformer | Particle Transformer (taus only) | Upcoming |
|
DeepTau v2.1 | CNN architecture (taus only) | TAU-20-001 |
A detailed comparison of different tau algorithms, accounting for differences in reconstruction efficiency can be found here: CMS-DP-2025-073.
The current recommemdation for Run 3 is DeepTau version 2.5 (DeepTau2018v2p5), scale factors for the UParT and PNet algorithms are being measured. Feedback from analysers is welcome!
Working Points for Different Algorithms
DeepTau v2.5
The DeepTau v2.5 algorithm is based on HPS taus and available in the tau collection for all Run 3 nanoAOD versions.
The working points (WPs) are included in central NanoAOD as idDeepTau2018v2p5VSjet (1 = VVVLoose, 2 = VVLoose, 3 = VLoose, 4 = Loose, 5 = Medium, 6 = Tight, 7 = VTight, 8 = VVTight), idDeepTau2018v2p5VSe (1 = VVVLoose, 2 = VVLoose, 3 = VLoose, 4 = Loose, 5 = Medium, 6 = Tight, 7 = VTight, 8 = VVTight) and idDeepTau2018v2p5VSmu (1 = VLoose, 2 = Loose, 3 = Medium, 4 = Tight).
The nanoAOD branches above should be used by most analyses, the raw scores corresponding to each of these WPs are shown here:
| Working Points | VVVLoose | VVLoose | VLoose | Loose | Medium | Tight | VTight | VVTight |
|---|---|---|---|---|---|---|---|---|
| VSjets | 0.3569 | 0.5910 | 0.7842 | 0.9167 | 0.9600 | 0.9781 | 0.9877 | 0.9927 |
| VSelectrons | 0.0917 | 0.1754 | 0.3410 | 0.6403 | 0.8359 | 0.9529 | 0.9829 | 0.9931 |
| VSmuons | N/A. | N/A. | 0.4432 | 0.7568 | 0.8825 | 0.9360 | N/A. | N/A. |
ParticleNet
The ParticleNet algorithm in the tau collection is based on CHS jets and available in the tau collection from nanoAOD v12.
In order to select a working point, a lower cut must be placed on the raw discriminator scores: rawPNetVSjet, rawPNetVSe and rawPNetVSmu.
The raw scores corresponding to each working point are:
| Working Points | VVVLoose | VVLoose | VLoose | Loose | Medium | Tight | VTight | VVTight |
|---|---|---|---|---|---|---|---|---|
| VSjets | 0.0565 | 0.1774 | 0.3810 | 0.6857 | 0.8347 | 0.9059 | 0.9494 | 0.9737 |
| VSelectrons | 0.1266 | 0.3547 | 0.6997 | 0.9354 | 0.9791 | 0.9897 | 0.9946 | 0.9971 |
| VSmuons | N/A. | N/A. | 0.2399 | 0.6037 | 0.8697 | 0.9451 | N/A. | N/A. |
UParT
The UParT algorithm in the tau collection is based on PUPPI jets and available in the tau collection from nanoAOD v14.
In order to select a working point, a lower cut must be placed on the raw discriminator scores: rawUParTVSjet, rawUParTVSe and rawUParTVSmu.
The raw scores corresponding to each working point are:
| Working Points | VVVLoose | VVLoose | VLoose | Loose | Medium | Tight | VTight | VVTight |
|---|---|---|---|---|---|---|---|---|
| VSjets | N/A. | N/A. | N/A. | 0.1316 | 0.2541 | 0.3956 | 0.5354 | 0.6649 |
| VSelectrons | 0.0688 | 0.2343 | 0.5691 | 0.8414 | 0.9455 | 0.9835 | 0.9931 | 0.9970 |
| VSmuons | N/A. | N/A. | 0.0025 | 0.0407 | 0.3492 | 0.8359 | N/A. | N/A. |
Certain VSjet working points aren't available in central NanoAOD, due to a cut of 0.05 that is applied as a baseline.
Summary of available tau algorithms (Run 2)
:right_arrow: Coming soon!