farbodtavakkoli/OTel-LLM-8.3B-Safety

OTel-LLM-8.3B-Safety: A Telecom-Specific Abstention Model

OTel-LLM-8.3B-Safety is an 8.3 billion parameter language model developed by farbodtavakkoli, specifically fine-tuned for the telecommunications sector. It is part of the broader OTel Family of Models, an initiative focused on creating open-source AI resources for global telecommunications. This particular variant is built upon the EssentialAI/rnj-1-instruct base model.

Key Capabilities & Differentiators

• Safety-Tuned Abstention: The model's core feature is its ability to abstain or refuse to answer when the provided context is insufficient or irrelevant to the user's query. This is crucial for maintaining accuracy and preventing hallucinations in RAG systems.
• Domain-Specific Training: It was fine-tuned using the OTel-Safety dataset, which comprises telecom-focused data curated by over 100 domain experts, including content from arXiv telecom papers, 3GPP standards, GSMA documents, and O-RAN specifications.
• RAG Pipeline Integration: Intended for use in telecom RAG pipelines where precise control over answer generation based on retrieved context is paramount.

Intended Use Cases

• Telecom RAG Systems: Ideal for applications requiring an assistant to explicitly state when it lacks sufficient information from retrieved documents.
• Reference for Safety Tuning: Can serve as a checkpoint for further specialized safety tuning efforts within telecommunications contexts.

Important Considerations

This model is an auxiliary safety variant and should be evaluated with abstention-focused metrics, such as correct-abstention rate on insufficient-context examples and answer quality on sufficient-context examples. It is domain-specific to telecommunications, English-only, and primarily text-centric, not a general-purpose LLM. Its performance is highly dependent on the quality of the retriever, reranker, context window, and prompt policy in the overall pipeline.