Development of Specialized Web-Search Subagent
A new specialized AI model, “Obsessive-1.7B,” has been released as a fine-tuned version of the LFM3-1.7B base model. According to reports on Reddit, the 1.1 GB model is designed to function specifically as a “subagent” for larger language models rather than as a standalone chat interface. Its primary objective is to process factual questions by executing a single web search and returning a concise answer consisting of a value, a verbatim quote, and a source URL.
The model was developed using LoRA rank 16 training on approximately 9,000 tool-calling conversations, utilizing an RTX 4060 laptop. The training data included Wikipedia and FineWeb pages, with a significant portion of the dataset dedicated to teaching the model to report when information is not found in search results.
Performance Benchmarks and Comparative Analysis
In a fixed-results benchmark involving 131 held-out SimpleQA questions, Obsessive-1.7B demonstrated significant improvements over its base model in specific metrics. According to the developer’s data, the subagent successfully found the answer in 94% of cases where the answer was present, compared to 50% for the base model. Notably, the subagent reported “not found” in 100% of cases where the answer was absent, whereas the base model failed to do so in 0% of those instances. Furthermore, the subagent showed a 0% rate of inventing answers, compared to 10% for the base model.
In live Wikipedia search tests involving 120 questions and 40 multi-fact requests, the results were more nuanced. While the subagent achieved a 100% return rate for any reply, it provided the correct answer in 24 out of 120 instances, trailing the base model’s 37 out of 120. However, the subagent successfully answered 40 out of 40 multi-fact requests, whereas the base model answered only one.
The developer noted that while the base model may produce more correct answers in live settings by continuing to search and write long replies, it fails to return any result 62% of the time because it runs out of turns.
Technical Limitations and Implementation
The developer has identified several known limitations of the Obsessive-1.7B model. The model is currently limited to one question per conversation and is trained exclusively in English. It has been observed to occasionally report claims as facts, quote neighboring facts incorrectly, or fail to find an answer that is present in the search results in approximately 10% of cases.
Technical discussions on Reddit suggest that the model’s small size may allow it to run on a CPU, potentially enabling it to operate in parallel with larger models without consuming significant GPU VRAM. However, some users raised concerns regarding “instruction injection,” where a model might mistake data within search results for instructions that modify its behavior.
The model is available via Hugging Face under the LFM Open License.