muradil211/AetherSearch_SFT
muradil211/AetherSearch_SFT is a 3.1 billion parameter Qwen2.5-3B-Instruct based causal language model fine-tuned by muradil211. It is specifically designed as a compact search agent, learning to reason, retrieve, and answer by generating search queries and integrating external information. This model excels at search-native behavior, including direct-answer, single-search, and multi-search trajectories, making it ideal for applications requiring agentic search and retrieval-augmented reasoning.
Loading preview...
AetherSearch SFT: A Compact Search Agent
AetherSearch SFT is a 3.1 billion parameter model, fine-tuned from Qwen2.5-3B-Instruct, designed to function as a compact search agent. It learns to reason, retrieve, and answer questions by interacting with an external search or RAG backend. The model was trained on 2,600 complete agent trajectories, encompassing direct-answer, single-search, and multi-search scenarios, with a context window of 32,768 tokens.
Key Capabilities
- Search-native behavior: The model intelligently determines when and what to search for, generating focused retrieval queries.
- Flexible search patterns: Trained on a diverse mix of 600 direct-answer, 1,025 single-search, and 975 multi-search examples, allowing it to adapt to various information retrieval needs.
- Evidence-in-the-loop reasoning: Integrates retrieved passages as context for reasoning while excluding them from training loss, ensuring grounded answers.
- Compact and accessible: Built on a 3B parameter backbone, making it suitable for accessible experimentation and deployment in resource-constrained environments.
- Reproducible: Includes public trainer, launcher, data checksum, and schema tests for transparent and verifiable results.
Good for
- Agentic search applications: Ideal for systems where the model needs to actively search for information to formulate answers.
- Retrieval-augmented generation (RAG): Designed to work seamlessly with external retrieval systems, where the host runtime executes search requests and provides information.
- Experimentation with search agents: Its compact size and reproducible training make it an excellent base for developing and testing new agentic reasoning strategies.
- Building custom QA services: Provides the core search-agent policy, requiring users to bring their own retriever and implement the XML-based retrieval loop.