TelecomTS overview: curation pipeline, covariates, and supported tasks.

ICML 2026 arXiv Hugging Face Dataset Python 3.11 License

TelecomTS is a large-scale, high-resolution, multi-modal dataset derived from a 5G telecommunications testbed. It is the first public observability dataset to preserve de-anonymized observability metrics with absolute scale information, encompassing by design a broad suite of multi-modal downstream tasks:

• 🔎 Anomaly detection (binary)
• 🛠️ Root-cause analysis (multi-class)
• ⏱️ Anomaly duration localization (sequence labeling)
• 📈 Forecasting / reconstruction (multi-channel)
• 🤖 Time series and network-level Q&A (multi-modal reasoning)

Observability data, particularly in telecommunications, differs fundamentally from conventional time series (e.g., weather, finance) by being:

• Zero-inflated, with metrics dominated by zeros punctuated by informative spikes
• Highly stochastic and bursty, with frequent, abrupt transitions
• Structurally noisy with minimal discernible temporal patterns

• ~32K time series samples and 1M+ total observations from a 5G testbed
• Multi-modal inputs:
  • Time series KPIs across PHY, MAC, and network layers, sampled at 10 Hz (100 ms)
  • Natural-language network descriptions and Q&A pairs
• Heterogeneous covariates: numeric KPIs and categorical fields (e.g., UL_Protocol, DL_Protocol)
• Absolute scale preserved (no normalization, no anonymization)
• Real and synthetic anomalies: 10 synthetic types grounded in telecom literature plus one real anomaly (jamming) collected over the air
• Reasoning traces: chain-of-thought traces for reasoning-aware fine-tuning and RL
• Labels / metadata: zone, application, mobility, congestion state, anomaly presence

Each sample in TelecomTS contains:

• start_time / end_time — temporal boundaries of the chunk

• sampling_rate_hz — number of timesteps per second

• description — natural-language summary of the network environment and time series behaviors

• KPIs — key performance indicator names and values

• anomalies — existence, type, duration, affected KPIs, and troubleshooting tickets

• statistics — mean, variance, trend, and periodicity for each KPI

• labels — contextual metadata (zone, application, mobility, congestion, anomaly presence)

• QnA — natural-language Q&A over the sample, grouped into timeseries, network, and anomalies subcategories. Each entry of has the following structure:

  { "q": "What activity was the user engaged in?",
   "a": "Twitch",
   "reasoning": "Sustained downlink throughput in the 2–4 Mbps range with periodic UDP bursts and stable RSRP is consistent with live video streaming..." }

  The reasoning field, present in the last two subcategories, contains an explicit reasoning trace that reveals the intermediate decision-making steps used to derive the final answer.

TelecomTS is hosted on the Hugging Face Hub at AliMaatouk/TelecomTS. You can load it directly with the 🤗 datasets library:

from datasets import load_dataset
dataset = load_dataset(
 "AliMaatouk/TelecomTS",
 data_files={"full": "**/chunked.jsonl"},
)["full"]
print(dataset) 

The benchmarking pipeline in this repo fetches the data automatically — no manual download is required.

Requires Python 3.11

# 1) Clone
git clone https://github.com/Ali-maatouk/TelecomTS.git
cd TelecomTS
# 2) Create & activate a virtual environment
python3.11 -m venv .venv
# macOS/Linux:
source .venv/bin/activate
# Windows (PowerShell):
# .venv\Scripts\Activate.ps1
# 3) Install dependencies
python -m pip install --upgrade pip
pip install -r requirements.txt
# 4) Run (uses configs/config.yaml)
# The dataset is fetched automatically from the Hugging Face Hub
# (AliMaatouk/TelecomTS) on first run and cached locally.
# This trains the selected encoder on the chosen task and then evaluates it.
python3 src/run.py

Choose the model and the task in configs/config.yaml. Running python3 src/run.py trains the selected model and evaluates it on the chosen task.

• Tasks (task_type)

  • anomaly detection
  • root-cause analysis
  • anomaly duration
  • forecasting

• Encoders (encoder_type)

  • TimesNet
  • Autoformer
  • NonStationary_Transformer
  • FEDformer
  • Informer

You can find the paper with all details at https://arxiv.org/abs/2510.06063. Please cite it as follows:

@misc{feng2025telecomtsmultimodalobservabilitydataset,
 title={TelecomTS: A Multi-Modal Observability Dataset for Time Series and Language Analysis},
 author={Austin Feng and Andreas Varvarigos and Ioannis Panitsas and Daniela Fernandez and Jinbiao Wei and Yuwei Guo and Jialin Chen and Ali Maatouk and Leandros Tassiulas and Rex Ying},
 year={2025},
 eprint={2510.06063},
 archivePrefix={arXiv},
 primaryClass={cs.AI},
 url={https://arxiv.org/abs/2510.06063},
}