OTel-Embedding: Contrastive Learning Dataset for Telecom Retrieval Models

OTel-Embedding is a contrastive learning dataset for training bi-encoder retrieval models using Multiple Negatives Ranking Loss. It is part of the Open Telco AI project, 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.

Use Cases

• Train bi-encoder retrieval models based on the described contrastive learning setup.
• Benchmark retrieval performance using the provided anchor queries and hard negative passages.
• Fine-tune models for telecom-specific information retrieval tasks based on domain expert curation.

Strengths

• Curated by over 100 domain experts from industry and academia.
• Designed for a specific training objective: Multiple Negatives Ranking Loss.
• Each sample includes one positive and five hard negative passages.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Open Telco (OTel) AI project.

Collection Method

Curated by domain experts.

Freshness

Last updated 2026-05-07 09:47:33; freshness should be verified.