TelecomTS: 5G Observability Dataset Benchmark
- TelecomTS is a de-anonymized, large-scale observability dataset featuring high-frequency, zero-inflated KPI streams from a 5G network.
- It supports anomaly detection, root-cause analysis, forecasting, and multi-modal QA by pairing diverse time series data with natural language annotations.
- Preserving absolute scale and heterogeneous covariates, TelecomTS provides critical operational insights that normalized benchmarks often overlook.
TelecomTS is a large-scale observability dataset derived from a 5G telecommunications network and designed for time series and language analysis (Feng et al., 7 Oct 2025). It targets a class of engineering data that differs sharply from conventional time series: observability data are described as zero-inflated, highly stochastic, and marked by minimal temporal structure, with abrupt, noisy, and high-variance dynamics. TelecomTS is presented as a de-anonymized, scale-preserving resource with heterogeneous covariates and native support for anomaly detection, root-cause analysis, forecasting, and multi-modal question answering, thereby addressing a gap left by proprietary restrictions and by public benchmarks that anonymize or normalize away operationally meaningful information (Feng et al., 7 Oct 2025).
1. Definition and motivation
TelecomTS is motivated by the claim that observability datasets are underrepresented in public benchmarks due to proprietary restrictions, while existing public datasets are often anonymized and normalized, removing scale information and limiting their use for tasks beyond forecasting (Feng et al., 7 Oct 2025). In this setting, observability data are metrics derived from monitoring complex engineering systems, including cloud and telecom networks, and they exhibit sparsity, zero-inflation, high stochasticity, abrupt dynamics, and minimal temporal structure.
A central premise of TelecomTS is that these properties make observability substantially different from the domains emphasized in standard time-series corpora such as weather, energy, or finance. The dataset is therefore positioned not merely as another forecasting benchmark, but as a resource for practical observability workloads in which variable identity, units, and absolute magnitudes remain intact. The paper characterizes it as the first large-scale, de-anonymized observability dataset derived from a real 5G telecom testbed and emphasizes that it is not anonymized or scale-normalized (Feng et al., 7 Oct 2025).
This positioning is significant because many operational tasks in telecom depend on the semantics of the monitored variables. A plausible implication is that preserving both the names and the scales of KPIs makes TelecomTS more suitable for deployment-oriented evaluation than benchmarks built around normalized, abstracted channels.
2. Dataset composition and measurement regime
TelecomTS contains over 1 million observations sampled at 100 ms intervals, corresponding to a 10 Hz sampling rate (Feng et al., 7 Oct 2025). It includes 18 channels or covariates: 10 floats, 6 integers, and 2 categorical variables. These channels span both device-side and base-station-side KPIs across all layers of the 5G protocol stack. The data were collected under static and mobile device conditions, across variable spatial zones defined by distance from the base station, and under different applications, including YouTube, Twitch, and file downloads. The collection also includes intentional adversarial conditions such as jamming, congestion, and multiple user profiles.
| Statistic | Value |
|---|---|
| Normal Observations | 1,020,000 |
| Anomalous Observations | 120,000 |
| Number of Channels | 18 (10 float, 6 int, 2 cat) |
| Sampling Rate | 10 Hz (100 ms) |
| Synthetic Anomaly Types | 10 |
| Jamming Observations | 30,000 |
| Traffic Types | YouTube, Twitch, File DL |
| QA Types (TS, Network) | 64, 5 |
The dataset is explicitly heterogeneous. It includes physical metrics such as RSRP and SNR, protocol and traffic statistics such as TX/RX bytes and protocol type, error indicators such as BLER, allocated resources, and categorical protocol state. The preserved scale information ranges from binary or categorical values to large-scale integers such as TX_Bytes in the interval (Feng et al., 7 Oct 2025).
The paper also emphasizes that many metrics are sparsely populated, with bursty abnormal events amid long runs of zeros. It states that fluctuations often do not represent anomalies and summarizes this regime with the observation that, in observability, erratic is the normal. This characterization is central to TelecomTS: the challenge is not only modeling variability, but distinguishing operationally normal irregularity from actual faults.
3. Task suite and multi-modal structure
TelecomTS supports four task families: anomaly detection, root-cause analysis, forecasting, and multi-modal question answering (Feng et al., 7 Oct 2025).
For anomaly detection, the dataset includes both real and synthetic anomalies. Real anomalies are induced by over-the-air jamming, which emulates adversarial interference events. Synthetic anomalies comprise 10 types curated using domain literature, with symptoms matching real fault modes such as sharp drops, oscillations, and faulty handover events. Their occurrences and durations are drawn from exponential distributions. Each anomalous segment is paired with annotated troubleshooting tickets generated and human-verified via GPT-4.1, which also makes the dataset suitable for root-cause analysis.
The multi-modal QA component is divided into time-series QA and network-context QA. Time-series QA covers questions about basic statistics such as mean and variance, as well as structure, including trend and periodicity, obtained via DFT and regression. Network-context QA covers application or user activity, congestion status, spatial location by zone, and mobility, extracted from metadata and rendered into natural-language QA pairs via GPT-4.1 (Feng et al., 7 Oct 2025).
This native pairing of time series and language is one of TelecomTS’s defining properties. The paper states that the dataset was designed to support joint time series-language analysis, rather than retrofitting text onto a pre-existing numerical corpus. This suggests a benchmark regime in which models must align KPI trajectories, operational context, and textual explanations rather than optimizing a single modality in isolation.
4. Benchmark results and characteristic failure modes
TelecomTS benchmarks LLMs, reasoning models, time-series foundation models, and established forecasting architectures, and reports that existing approaches struggle with the abrupt, noisy, and high-variance dynamics of observability data (Feng et al., 7 Oct 2025). The evaluated systems include GPT-4.1, Claude 3.7 Sonnet, o4-mini, DeepSeek-R1, Moment, Moirai2, Toto, Mantis, TimesNet, Autoformer, FEDformer, Informer, and the Non-stationary Transformer.
Representative reported results illustrate the difficulty of the dataset:
| Task | Representative reported values | Observed pattern |
|---|---|---|
| Anomaly detection | GPT-4.1 without context: Precision 0.054, Recall 1.000, F1 0.103; Mantis: Precision 0.640, Recall 0.800, F1 0.711 | LLMs show very high recall but extremely low precision |
| Root-cause analysis | GPT-4.1 with context: 0.23 accuracy; o4-mini with context: 0.28; Toto: 0.85 | LLMs are near random; observability-pretrained TS models perform better |
| Forecasting | Informer: MAE 0.1437, RMSE 0.3586; FEDformer: 0.1702, 0.4080; Mantis: 0.4578, 0.6037 | Reasonable aggregate errors coexist with failure on abrupt bursts and peaks |
| Multi-modal QA | Network QA accuracy: GPT-4.1 53.3, Claude 95.0, o4-mini 76.7, DeepSeek-R1 98.3 | Performance varies widely across modalities and question types |
The anomaly-detection results are especially revealing. The paper states that language and reasoning models tend to cry wolf: they achieve very high recall but extremely low precision because they interpret normal erratic spikes as anomalies. Even a normal burst in TX_Bytes, such as a user streaming video, often triggers false positives across model classes. By contrast, anomaly-boundary estimation becomes markedly easier once the presence of an anomaly is already known, with Mantis reaching an F1 of approximately 0.89 for anomaly duration (Feng et al., 7 Oct 2025). The main difficulty is therefore anomaly existence, not merely localization.
In root-cause analysis, LLMs remain close to random even when contextual text is supplied, whereas time-series models with observability pretraining perform substantially better. The strongest reported value in the summary is Toto at 0.85 accuracy. In forecasting, the paper warns that pointwise metrics can be misleadingly optimistic because long regions of stability dominate the error even though models fail on the abrupt bursts and peaks that matter operationally.
5. Absolute scale as a modeling principle
A major contribution of TelecomTS is the empirical argument that absolute scale must be preserved for practical observability modeling (Feng et al., 7 Oct 2025). The paper contrasts this with a common preprocessing step used by many architectures:
According to the benchmark analysis, this kind of pre-normalization destroys scale-dependent features that are critical for anomaly detection and root-cause analysis. The paper states that models relying on normalized or zero-meaned data cannot avoid high false positives or missed anomalies, whereas methods that process the original scale or explicitly embed scale information perform much better. Mantis is singled out as an example of a model that augments its representation with patch mean and standard deviation and correspondingly performs more robustly (Feng et al., 7 Oct 2025).
The underlying argument is operational rather than purely statistical. A CPU load of 0.7, a BLER excursion, a packet-drop count, or a signal-strength value have direct meaning because their absolute magnitudes and thresholds matter. Losing that information by normalization weakens not only forecasting but also diagnosis and causal interpretation. The paper therefore concludes that foundation time series models should natively leverage scale information for practical observability applications.
This emphasis distinguishes TelecomTS from benchmarks built primarily for normalized forecasting tasks. A plausible implication is that TelecomTS can serve as a stress test for architectures whose inductive biases were developed on smoother, more stationary datasets.
6. Relation to adjacent telecom benchmarks and common confusions
TelecomTS occupies a specific place within a broader telecom-data landscape. Other telecom resources summarized in the literature focus on different modalities and evaluation problems. For example, “Understanding Telecom Language Through LLMs” adapts BERT, DistilBERT, RoBERTa, and GPT-2 to telecom language using 3GPP technical documents and evaluates working-group classification, with BERT and RoBERTa reaching 84.6% accuracy (Bariah et al., 2023). “TeleEmbedBench” is a multi-corpus embedding benchmark for retrieval-augmented generation in telecommunications, spanning O-RAN Alliance specifications, 3GPP release documents, and the srsRAN codebase, with 9,000 question-chunk pairs across three chunk sizes (Gajjar et al., 20 Apr 2026). “Efficient Telecom Specific LLM: TSLAM-Mini with QLoRA and Digital Twin Data” focuses on instruction-following and response quality for telecom use-cases using a 100,000-sample domain dataset and automated LLM-based judging (Ethiraj et al., 10 May 2025).
These works indicate that telecom AI evaluation has largely centered on standards text, retrieval, and telecom-specific language modeling. This suggests that TelecomTS fills a different niche: observability-centric, scale-preserving, and explicitly multi-modal across KPIs and natural-language explanations.
A recurrent source of confusion is the similarity between TelecomTS and TelecomTM. They are unrelated artifacts. TelecomTM is a traffic monitoring system that uses pre-existing telecommunication fiber-optic cables as virtual strain sensors for vehicle detection, tracking, and characterization, achieving 90.18% vehicle detection accuracy in a real-world evaluation (Liu et al., 2023). TelecomTS, by contrast, is a 5G observability dataset for time series and language analysis (Feng et al., 7 Oct 2025).
Taken together, the available evidence positions TelecomTS as a benchmark for a class of telecom problems that conventional time-series and LLM benchmarks do not capture well: zero-inflated, high-variance, mixed-type, de-anonymized observability streams with paired language supervision. Its reported results support the stronger conclusion that current state-of-the-art models remain poorly matched to this regime and that progress will likely require scale-aware, modality-flexible foundation models rather than straightforward transfer from classic forecasting or document-understanding settings (Feng et al., 7 Oct 2025).