farbodtavakkoli/OTel-LLM-12B-Safety

Hugging Face
VISIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kPublished:Feb 26, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

OTel-LLM-12B-Safety by farbodtavakkoli is a 12 billion parameter, safety-tuned telecom language model, fine-tuned on OTel telecommunications data. Based on google/gemma-3-12b-it, this model is specifically designed to refuse answers when retrieved context is insufficient or off-topic. It excels in telecom RAG pipelines where abstention behavior is critical for reliability and accuracy.

Loading preview...

Overview

OTel-LLM-12B-Safety is a 12 billion parameter language model developed by farbodtavakkoli, specifically fine-tuned for the telecommunications sector. It is built upon the google/gemma-3-12b-it base model and is part of the OTel Family of Models, an initiative to provide open-source AI resources for telecom. This particular variant is safety-tuned to exhibit abstention behavior, meaning it is trained to refuse to answer when the provided context is insufficient or irrelevant.

Key Capabilities

  • Safety-tuned Abstention: Designed to abstain from answering when retrieved context is missing, irrelevant, or insufficient, crucial for reliable RAG applications.
  • Telecom Domain Expertise: Fine-tuned on a large dataset of telecommunications data, including arXiv papers, 3GPP standards, GSMA documents, and O-RAN specifications, curated by over 100 domain experts.
  • Full-Parameter Fine-tuning: Underwent full-parameter post-training on the OTel-Safety dataset, which focuses on examples requiring abstention.

Good For

  • Telecom RAG Pipelines: Ideal for use in Retrieval Augmented Generation (RAG) systems within the telecommunications industry where accurate and context-aware responses are paramount.
  • Abstention-Focused Generation: Suitable for scenarios requiring a model to explicitly state when it cannot provide a confident answer based on available information.
  • Reference Checkpoint: Can serve as a base for further safety tuning or as a component in complex AI systems requiring robust refusal mechanisms.