README.md

---
license: mit
configs:
- config_name: troubleshooting
  data_files:
  - split: test
    path: "troubleshooting/test.json"
  - split: train
    path: "troubleshooting/train.json"
extra_gated_fields:
  First Name: text
  Last Name: text
  Affiliation: text
  Country: country
  Job title:
    type: select
    options:
      - Student
      - Research Graduate
      - Researcher
      - Engineer
      - Reporter
      - Other
    geo: ip_location
---
<div align="center" style="font-family: 'Helvetica Neue', Helvetica, Arial, sans-serif; padding: 25px 15px; max-width: 720px; margin: 20px auto;">
  <div style="font-size: 3.2em; font-weight: bold; margin-bottom: 5px;">
    <span style="background: -webkit-linear-gradient(45deg,rgb(247, 233, 43),rgb(98, 10, 61)); -webkit-background-clip: text; -webkit-text-fill-color: transparent;">
      TeleLogs
    </span>
  </div>
  <div style="font-size: 1.0em; color: #4a4a4a; margin-bottom: 12px; line-height: 1.45; padding: 0 10px;">
    A Benchmark for Large Language Models in Telecom Data Analysis
  </div>
  <div style="font-size: 0.80em; color: #777; margin-bottom: 10px;">
    Developed by the <strong>NetOp Team, Huawei Paris Research Center</strong>
  </div>
  <hr style="border: 0; height: 1px; background: #ddd; margin-top: 10px; margin-bottom: 15px; width: 60%;">

  <div style="display: flex; flex-wrap: wrap; justify-content: center; gap: 10px; font-size: 1.1em; margin-bottom: 0px;">
  <a href="https://arxiv.org/abs/2507.21974" target="_blank"
     style="text-decoration: none; background-color: #007bff; color: white; padding: 10px 20px; border-radius: 5px; font-weight: bold; text-align: center;">
    📄 Read the Paper
  </a>
  <a href="https://huggingface.co/datasets/netop/TeleLogs"
     style="text-decoration: none; background-color: #ffc107; color: black; padding: 10px 20px; border-radius: 5px; font-weight: bold; text-align: center;">
    🤗 Explore the Dataset
  </a>
  </div>
</div>


> [!NOTE]
> IMPORTANT: Please help us protect the integrity of this benchmark by not publicly sharing, re-uploading, or distributing the dataset.


## Dataset Description

- **Repository (Dataset & Evaluation Code):** [https://huggingface.co/datasets/netop/TeleLogs](https://huggingface.co/datasets/netop/TeleLogs)
- **Paper:** [https://arxiv.org/abs/2507.21974](https://arxiv.org/abs/2507.21974)

**TeleLogs** is a synthetic dataset designed to advance research on **Root Cause Analysis (RCA)** in 5G networks.
Unlike general-purpose troubleshooting datasets, TeleLogs is built on **realistic network engineering parameters** and simulates drive-test scenarios involving a user equipment (UE) moving through a region covered by multiple 5G base stations (gNodeBs).
The dataset provides **fine-grained observability of network configuration and user-plane performance**, enabling the systematic study of faults such as misconfigured mobility parameters, antenna misalignment, or interference. 
TeleLogs includes both user-plane measurements (e.g., throughput, RSRP, SINR) and network configuration data (e.g., antenna azimuth, downtilt, and resource allocation).

Each instance in TeleLogs involves:
* A **symptom** (throughput degradation below 600 Mbps).
* A set of **root causes** (one or more among 8 predefined causes).
* A **full context of network and mobility features** for diagnosis.

The dataset is released with both **training and test splits** to facilitate reproducibility and benchmarking of RCA models.

## Dataset format
### Network Engineering Parameters
Each gNodeB is described by detailed configuration attributes:
* Physical deployment: location (latitude, longitude), height, and cell IDs.
* Antenna settings: azimuth, downtilt, transmit/receive mode.
* Beamforming scenarios and co-frequency neighbor cell information.

### User-Plane Data
Collected during the drive test simulation, user-plane records include:
* Timestamp and vehicle speed.
* Downlink throughput.
* Serving cell metrics such as reference signal received power (RSRP) and signal-to-interference-plus-noise ratio (SINR).
* Top-k neighbor cell RSRPs and phisical cell IDs (PCIs).
* Resource block allocation.

### Observed Symptom
The diagnostic focus is on downlink throughput degradation, defined as throughput < 600 Mbps.

### Root Causes
TeleLogs includes **8 root causes**, each reflecting a specific misconfiguration or network issue:
1. Test vehicle speed exceeds 40 km/h, impacting user throughput.
2. The downtilt angle of the serving cell is too large, causing weak coverage at the far end.
3. The serving cell coverage distance exceeds 1 km, resulting in poor RSRP.
4. Non-colocated co-frequency neighboring cells cause severe interference.
5. Neighbor cell and serving cell have the same physical cell ID (PCI) mod 30. As a result, their reference signals can overlap, leading to interference.
6. Frequent handovers degrading user performance.
7. Misconfigured handover thresholds degrading user performance.
8. The average scheduled resource blocks (RBs) of the serving cell are below 160, which affects user throughput


## LLM Performance on TeleLogs
We benchmarked both **state-of-the-art (SoTA) LLMs** and our **fine-tuned Qwen2.5-RCA models (1.5B, 7B, 32B)** on the TeleLogs dataset. Performance was evaluated using:

* **pass@1**: Accuracy of producing a correct root cause analysis in a single attempt.
* **maj@4**: Accuracy under majority voting across 4 independent attempts.

Our fine-tuned models dramatically outperform both general-purpose instruction-tuned models and reasoning-oriented LLMs. 
The **Qwen2.5-RCA-32B** achieves **95.86% pass@1** and **96.18% maj@4** on standard TeleLogs test, surpassing the strongest baseline (Qwen3-32B) by **nearly 3×**.

| Model                             | Reasoning | Test pass@1 | Test maj@4 | Randomized pass@1 | Randomized maj@4 |
| :-------------------------------  | :-------: | -----------:| ----------:| -----------------:| -----------------:|
| Qwen2.5-1.5B-Instruct         | ❌         | 11.25%      | 11.60%     | 9.15%             | 9.80%             |
| Qwen2.5-7B-Instruct           | ❌         | 12.05%      | 10.80%     | 11.50%            | 11.80%            |
| Qwen2.5-32B-Instruct          | ❌         | 18.85%      | 19.60%     | 18.05%            | 18.70%            |
| DeepSeek-R1-Distill-Llama-70B     | ✅         | 29.42%      | 34.84%     | 29.00%            | 32.18%            |
| QwQ-32B                           | ✅         | 33.62%      | 39.00%     | 32.14%            | 38.86%            |
| Qwen3-32B                         | ✅         | 33.77%      | 37.04%     | 31.37%            | 36.23%            |
| **Qwen2.5-RCA-1.5B**              | ✅         | **87.56%**  | **87.73%** | **75.90%**        | **77.08%**        |
| **Qwen2.5-RCA-7B**                | ✅         | **87.01%**  | **88.89%** | **77.95%**        | **80.32%**        |
| **Qwen2.5-RCA-32B**               | ✅         | **95.86%**  | **96.18%** | **93.23%**        | **95.02%**        |

## Citation

If you use TeleLogs in your research, please cite our paper.
```bibtex
@article{sana2025reasoning,
  title={{Reasoning Language Models for Root Cause Analysis in 5G Wireless Networks}},
  author={Mohamed Sana and Nicola Piovesan and Antonio De Domenico and Yibin Kang and Haozhe Zhang and Merouane Debbah and Fadhel Ayed},
  year={2025},
  eprint={[arXiv preprint arXiv:2507.21974]},
  url={https://arxiv.org/abs/2507.21974}
}