Omokemi/hop-1-gemma-270m

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.3BQuant:BF16Context Size:32kPublished:Aug 20, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

Omokemi/hop-1-gemma-270m is a 0.3 billion parameter Gemma-based model fine-tuned by Omokemi for multi-hop question answering. It specializes in identifying the first lookup step for complex questions, outputting a YAML structure with thought, action, and target entity. This model is optimized for offline, cost-effective initial reasoning steps in multi-hop systems, rather than providing final answers.

Loading preview...

Model Overview

Omokemi/hop-1-gemma-270m is a specialized 270-million-parameter model, fine-tuned from Google's Gemma-3-270M-it, designed for the initial step of multi-hop question answering. Its core function is to take a complex question and output the first piece of information or entity that needs to be looked up, formatted as YAML with thought, action, and target_entity fields. It does not perform document retrieval or provide the final answer to the question.

Key Capabilities & Differentiators

  • Specialized Multi-Hop Reasoning: Focuses exclusively on decomposing multi-hop questions into their first logical lookup step, a unique niche for smaller models.
  • Cost-Effective & Offline Operation: Unlike larger, API-dependent models, this model can run entirely offline, incurring no per-call costs once loaded, making it suitable for privacy-sensitive or resource-constrained environments.
  • Strong Format Adherence: Achieves 100% format adherence and 87.4% target correctness on its specialized task when using a fixed system prompt, outperforming the untrained base model significantly.
  • Efficient for its Size: While not outperforming much larger models in accuracy, it demonstrates strong performance for its compact size, showing that fine-tuning can enable specific reasoning tasks on smaller architectures.

Use Cases

  • Pre-processing for RAG Systems: Ideal as a lightweight, initial component in a larger Retrieval-Augmented Generation (RAG) pipeline to guide information retrieval.
  • Offline Reasoning Agents: Suitable for applications requiring local, network-independent decomposition of complex queries.
  • Educational Tools: Can be used to demonstrate multi-hop reasoning decomposition without relying on expensive external APIs.