Obsessive-1.7B:1.1 GB 网页搜索子智能体
一名 Reddit 作者以 LoRA rank 16、使用约 9,000 条工具调用对话微调 LFM3-1.7B,使其专注于一项任务:接收事实性问题,调用一次 web_search 后返回简短答案、逐字引用和 URL;若结果中没有答案,则回答“Not found in the search results.”。该模型以 1.1 GB GGUF 文件形式在 Hugging Face 发布,沿用基础模型的 LFM Open License,并提供 Pi、OpenCode 和 Hermes 的配置及 llama-server 运行命令。作者自测显示,该模型在答案存在时找到答案的比例为 50%,编造答案占 10%,在 RTX 4060 笔记本上每题中位响应时间为 5.1 秒。在使用未见过的后端进行实时 Wikipedia 搜索时,它在 120 道题中答对 37 道,并完成 40 个多事实请求中的 1 个。作者还称,该模型会把结果中实际存在的答案误报为“未找到”的比例约为十分之一,仅支持英语,并适合每次对话处理一个问题。
此次发布提供了一个可在更大模型下运行、专注于单次网页搜索的小型可运行模型。作者自报的评测同时显示其存在编造答案、漏掉答案和多事实请求表现不佳等明显局限,因此适用性取决于具体部署方案。
I fine-tuned LFM3-1.7B to do one job: take a factual question, call web_search once, and answer with a short value, a word-for-word quote and the URL. If the results don’t contain the answer, it says “Not found in the search results.” It’s meant to run as a subagent under a bigger model, not as a chat model. I have to say the base lfm really surprised me at its capability to run tool calls and extract information.
Answer: Barbara Liskov Evidence: "Liskov ... the second woman to receive the Turing Award." Source: https://en.wikipedia.org/wiki/Barbara_Liskov Fixed-results benchmark (131 held-out SimpleQA questions; in 29 of them the answer was removed from the results):
| base | Obsessive |
|---|---|
| Finds the answer when it’s there | 50% |
| Says “not found” when it isn’t | 0% |
| Invented answers | 10% |
| Median sec/question (4060 laptop) | 5.1 |
#Obsessive-1.7B#LFM3-1.7B#Liquid AI#llama.cpp#Hugging Face#Qwen 3.8 Flash Next