Supreeth/searchlm-nl2bm25-sft
Supreeth/searchlm-nl2bm25-sft is a 3.1 billion parameter Qwen2.5-3B-Instruct model, fine-tuned by Supreeth via LoRA SFT to convert natural language queries into Tantivy boolean search queries. This model generates explicit chain-of-thought reasoning alongside the structured query output. It serves as a warm-start checkpoint for the SearchLM collection, specifically designed for robust information retrieval by translating user intent into precise search engine syntax.
Loading preview...
Overview
Supreeth/searchlm-nl2bm25-sft is a 3.1 billion parameter model, part of the SearchLM collection, specifically fine-tuned from Qwen2.5-3B-Instruct. Its core function is to transform natural language queries into structured Tantivy boolean search queries, complete with a detailed chain-of-thought reasoning. This model represents the warm-start checkpoint (SFT v1) in a multi-stage training pipeline, preceding more advanced GRPO (Generative Reinforcement Learning with Policy Optimization) versions like GRPO v2.
Key Capabilities
- Natural Language to Boolean Query Conversion: Translates user queries like "Do statins cause breast cancer?" into precise Tantivy boolean syntax.
- Chain-of-Thought Reasoning: Provides an explicit
<reasoning>block detailing concept extraction, synonym expansion, and query strategy. - Structured Output: Generates a
<query>block with valid Tantivy syntax, ready for direct use with search engines. - LoRA SFT Fine-tuning: Utilizes LoRA (r=16, α=32) on a Qwen2.5-3B-Instruct base, targeting q/k/v/o and gate/up/down projections.
- Training Data: Trained on 4,999 examples from the Supreeth/nl2bm25-sft dataset, derived from BEIR datasets (NFCorpus, SciFact, FiQA-2018, ArguAna, HotpotQA, NQ).
Good For
- Developers building search applications that require precise boolean query generation from natural language inputs.
- Integrating advanced information retrieval capabilities into systems using the Tantivy search engine.
- As a foundational model for further fine-tuning or research in query generation, particularly for understanding the initial SFT stage before reinforcement learning.