~/FINE TUNING/obsessive-1-7b-a-1-1-gb-web-search-subagent-for-local

Obsessive-1.7B: A 1.1 GB Web-Search Subagent for Local LLMs

Developer MuzafferMahi released Obsessive-1.7B, a 1.1 GB fine-tuned subagent built on LFM3-1.7B designed specifically to execute a single web search, extract precise facts with verbatim quotes, and reliably handle negative cases. It scores 94% on identifying unanswerable questions on a held-out SimpleQA benchmark, a massive improvement over the base model's 0%. This release demonstrates a practical, highly efficient approach to fine-tuning small language models for hyper-specific tool-use roles rather than general chat. By acting as a specialized subagent under a larger orchestrating model, it reduces hallucinations and prevents wasted turns on unanswerable queries. Trained using LoRA rank 16 on roughly 9,000 tool-calling conversations, Obsessive-1.7B is restricted to single-turn conversations and English only. Known limitations include occasionally reporting claims as absolute facts and occasionally quoting neighboring facts.

## BACKGROUND

Small Language Models (SLMs) are compact AI models that can run efficiently on consumer hardware like laptops or edge devices. Tool use, or function calling, allows an LLM to request external tools such as web searches, processing the returned data to augment its capabilities.

## REFERENCES

## KEYWORDS

#Fine-tuning#AI Agents#Local LLM#Tool Use#Small Language Models

$ subscribe --daily