OTel-LLM-8.3B-Classification

OTel-LLM-8.3B-Classification is a telecom-specialized classification model fine-tuned for 5G network root cause analysis (RCA), achieving ~99% accuracy on the TeleLogs benchmark. It is part of the OTel Family of Models, an open-source initiative to build industry-standard AI models for the global telecommunications sector.

Model Details

Attribute	Value
Base Model	EssentialAI/rnj-1
Parameters	8.3B
Training Method	Full parameter fine-tuning with classification head
Task	Multi-class sequence classification (8 root cause classes)
Language	English
License	Apache 2.0

Benchmark Results

This model achieves ~99% accuracy on the TeleLogs test set, higher than the published baselines in the TeleLogs paper shown below, including the reasoning-augmented models. It is reported as an auxiliary experiment in the OTel paper (NeurIPS 2026) and is not part of the paper's 30-model baseline table.

Model	Reasoning	Test pass@1	Test maj@4
Qwen2.5-32B-Instruct	❌	18.85%	19.60%
DeepSeek-R1-Distill-Llama-70B	✅	29.42%	34.84%
QwQ-32B	✅	33.62%	39.00%
Qwen3-32B	✅	33.77%	37.04%
Qwen2.5-RCA-1.5B	✅	87.56%	87.73%
Qwen2.5-RCA-7B	✅	87.01%	88.89%
Qwen2.5-RCA-32B	✅	95.86%	96.18%
OTel-LLM-8.3B-Classification	❌	~99%	—

TeleLogs Dataset

TeleLogs is a synthetic dataset designed to advance research on root cause analysis in 5G networks. It simulates drive-test scenarios involving a user equipment (UE) moving through a region covered by multiple 5G base stations (gNodeBs). Each instance includes a symptom (throughput degradation below 600 Mbps) and one or more root causes from 8 predefined classes:

Test vehicle speed exceeds 40 km/h, impacting user throughput
Downtilt angle of the serving cell is too large, causing weak coverage at the far end
Serving cell coverage distance exceeds 1 km, resulting in poor RSRP
Non-colocated co-frequency neighboring cells cause severe interference
Neighbor cell and serving cell have the same PCI mod 30, causing reference signal overlap
Frequent handovers degrading user performance
Misconfigured handover thresholds degrading user performance
Average scheduled resource blocks (RBs) of the serving cell are below 160

For more details, see the TeleLogs paper.

Training Details

Approach

The model was trained as a sequence classification model. The classification head consists of a Dropout(0.1) layer followed by a Linear projection from the model's hidden dimension to the 8 root cause classes.

Hyperparameters

Parameter	Value
Learning Rate	7e-6
LR Scheduler	Cosine
Warmup Steps	150
Batch Size (per device)	4
Gradient Accumulation Steps	2
Effective Batch Size	64 (4 × 2 × 8 GPUs)
Epochs	25
Weight Decay	0.01
Label Smoothing	0.1
Max Gradient Norm	0.5
Max Sequence Length	5000 tokens
Precision	BF16
Attention	Flash Attention 2
Distributed Training	FSDP
Gradient Checkpointing	Enabled

Classification Head

Dropout(0.1) → Linear(hidden_size, 8, bias=False)

Weight initialization: Normal distribution with mean=0.0, std=0.01.

Intended Use

This model is optimized for:

5G network root cause analysis — classifying network log data into specific fault categories
Telecom troubleshooting — automated diagnosis of throughput degradation causes from drive-test measurements

Related Models

Language Models

Embedding Models

OTel Embedding Collection

Reranker Models

OTel Reranker Collection

Related Datasets

Training Infrastructure

Framework: ScalarLM (GPU-agnostic)
Compute: TensorWave with AMD GPUs and Azure with NVIDIA GPUs
Training Code: OTel code

Resources

Citation

@inproceedings{tavakkoli2026otel,
  title     = {OTel: Open Telco AI Datasets, Benchmarks, and Models},
  author    = {Tavakkoli, Farbod and others},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026},
  url       = {https://github.com/farbodtavakkoli/OTel}
}

If you use the TeleLogs dataset, please also cite:

@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:2507.21974},
  url={https://arxiv.org/abs/2507.21974}
}

Contact

If you have any technical questions, please feel free to reach out to farbod.tavakkoli@att.com or farbodtavakoli@gmail.com

farbodtavakkoli/OTel-LLM-8.3B-Classification