File size: 5,230 Bytes
4805cd0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 | ---
license: apache-2.0
task_categories:
- sentence-similarity
- feature-extraction
language:
- en
tags:
- telecommunications
- telecom
- 3gpp
- oran
- gsma
- rag
- embedding
- retrieval
- bi-encoder
- mnrl
- domain-specific
- otel
size_categories:
- 100K<n<1M
pretty_name: OTel-Embedding
dataset_info:
features:
- name: anchor
dtype: string
- name: positive
dtype: string
- name: negative_1
dtype: string
- name: negative_2
dtype: string
- name: negative_3
dtype: string
- name: negative_4
dtype: string
- name: negative_5
dtype: string
---
# OTel-Embedding
## Dataset Summary
OTel-Embedding is a contrastive learning dataset for training bi-encoder retrieval models using Multiple Negatives Ranking Loss (MNRL). It is part of the Open Telco (OTel) AI project, the largest open-source AI initiative in telecommunications, curated by over 100 domain experts from industry and academia.
Each sample consists of an anchor query, one positive passage, and five hard negative passages. The dataset is designed to train embedding models that can accurately retrieve relevant telecom documentation given a natural language query, forming the retrieval component of end-to-end RAG pipelines in telecommunications.
## Dataset Structure
Each sample contains the following fields:
| Field | Type | Description |
|---|---|---|
| `anchor` | string | The query or information need used as the anchor in contrastive learning. |
| `positive` | string | A passage that is relevant to the anchor query. |
| `negative_1` | string | A hard negative passage that is topically related but does not answer the anchor query. |
| `negative_2` | string | A hard negative passage. |
| `negative_3` | string | A hard negative passage. |
| `negative_4` | string | A hard negative passage. |
| `negative_5` | string | A hard negative passage. |
The five hard negatives per sample are selected to be semantically proximate to the anchor but factually non-relevant, forcing the model to learn fine-grained distinctions within the telecom domain.
## Data Sources and Contributor Attribution
The OTel corpus was curated by 100+ domain experts from industry and academia. Each source class was contributed by a specific institutional partner:
| Source | Contributor | Raw samples |
|---|---|---:|
| arXiv telecom papers, 3GPP standards, telecom Wikipedia, telecom Common Crawl pages | Yale University | 681,172 |
| GSMA Permanent Reference Documents (PRDs), Discover portal, mixed telecom documents | GSMA | 158,006 |
| IETF RFC series | NetoAI | 100,751 |
| Industry whitepapers | Khalifa University | 62,000 |
| O-RAN specifications across working groups 1, 2, 4, 5, 6, 7, 8, 9, 10 | University of Leeds | 58,565 |
| O-RAN documents across working groups | The University of Texas at Dallas | 42,000 |
| **Total raw samples** | | **~1,102,494** |
After the four-stage cleaning pipeline (described below), the corpus was reduced to **326,767** higher-confidence examples across the OTel dataset family.
## Data Collection and Processing
The raw corpus comprised approximately 1.1 million samples. A four-stage cleaning pipeline reduced this to 326K samples across the OTel dataset family:
1. **Heuristic filtering**: Rule-based removal of malformed, duplicated, or low-quality entries.
2. **Reranking filtering**: Relevance scoring with a cross-encoder reranker to discard weakly aligned query-passage pairs.
3. **Embedding filtering**: Semantic quality assessment with a strong embedding model to remove outliers and near-duplicates in embedding space.
4. **Deduplication**: Final deduplication pass to ensure sample uniqueness.
Hard negatives were mined from the corpus using dense retrieval to identify passages that are close in embedding space to the positive but do not satisfy the information need of the anchor.
## Intended Use
OTel-Embedding is intended for training bi-encoder embedding models for telecom-domain retrieval. Target applications include:
- Fine-tuning sentence transformer models for telecom document retrieval using MNRL or similar contrastive objectives.
- Building the retrieval stage of RAG pipelines over telecom standards and specifications.
- Benchmarking domain-adapted embedding models against general-purpose alternatives on telecom retrieval tasks.
## Limitations
- **Domain scope**: The dataset covers telecommunications standards and specifications. Embedding models trained on this data are optimized for telecom retrieval and may not generalize well to other domains.
- **Language**: English only.
- **Negative quality**: Hard negatives are mined automatically and may include occasional false negatives (passages that are in fact relevant to the anchor).
- **Temporal coverage**: The source documents reflect standards and specifications available at the time of collection. Newer revisions may not be represented.
## Citation
```bibtex
@misc{otel_data_2026,
title = {{OTel}: Open Telco AI Datasets, Benchmarks, and Models},
author = {Tavakkoli, Farbod and others},
year = {2026},
note = {Open Telco (OTel) dataset release},
url = {https://huggingface.co/farbodtavakkoli}
}
``` |