farbodtavakkoli/OTel-LLM-20B-Reasoning

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:2Model Size:20BQuant:FP8Context Size:128kPublished:Feb 15, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

The farbodtavakkoli/OTel-LLM-20B-Reasoning is a 20 billion parameter, full-parameter fine-tuned language model developed by Farbod Tavakkoli, based on openai/gpt-oss-20b. It is specifically optimized for context-grounded reasoning within the telecommunications domain, utilizing the OTel-LLM dataset. This model excels at generating accurate answers to telecom-related questions when provided with relevant context, making it suitable for Retrieval-Augmented Generation (RAG) pipelines in the telecom sector.

Loading preview...

OTel-LLM-20B-Reasoning: A Specialized Telecom LLM

OTel-LLM-20B-Reasoning is a 20 billion parameter language model developed by Farbod Tavakkoli, fine-tuned from openai/gpt-oss-20b specifically for the telecommunications sector. It is part of the OTel Family of Models, an open-source initiative to provide AI resources for global telecommunications.

Key Capabilities & Features

  • Domain-Specific Expertise: Full-parameter fine-tuned on the OTel-LLM dataset, which comprises telecom-focused data curated by over 100 domain experts.
  • Context-Grounded Reasoning: Designed for Retrieval-Augmented Generation (RAG) pipelines, improving context-grounded correctness by +3.7 to +10.0 percentage points over base checkpoints.
  • Performance: Achieves 71.7% LLM-as-judge correctness on held-out OTel-LLM evaluation data, a +6.5 pp improvement over its base model's 65.2%.
  • Training Details: Trained using ScalarLM, AdamW optimizer, BF16 precision, and Flash Attention 2, with a maximum sequence length of 1500 tokens.

Intended Use Cases

  • Telecom Answer Generation: Ideal for generating answers to telecom-specific questions when provided with retrieved context.
  • RAG Pipelines: Best utilized within RAG architectures where external telecom knowledge is supplied.

Limitations

  • Domain-Specific: Not intended as a general-purpose language model.
  • English-Only: Primarily text-centric and currently supports only the English language.
  • Context-Dependent: Performance is optimized for context-grounded tasks; not suitable for unrestricted context-free QA without separate evaluation.

Finetunes
1 models