# Obsessive-1.7B: a 1.1 GB web-search subagent — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/b79a45d90bff2f5c

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-07 12:13 ingested 2026-11-08 02:39

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xq4yju/obsessive17b_a_11_gb_websearch_subagent/)

## retained text



```
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%|94%|
|Says “not found” when it isn’t|0%|100%|
|Invented answers|10%|0%|
|Median sec/question (4060 laptop)|5.1|2.0|

**Live Wikipedia search** (120 questions + 40 multi-fact requests, a search backend it never trained on):

|base|Obsessive|
|:-|:-|
|Returns any reply|38%|100%|
|Correct answer in reply|**37/120**|24/120|
|Multi-fact requests answered|1/40|40/40|
|Invented fact values|4 of 4|3 of 52|

The base model wins on raw correct answers live: it keeps searching and writes long replies, so the answer turns up more often, but 62% of the time it runs out of turns and returns nothing. Obsessive searches once and only states what it can quote. Whether that trade suits you depends on your setup.

**Known problems**

* One question per conversation. Multi-turn chats confuse it (only trained on single questions, limited VRAM for training :/ ).
* Reports claims as facts: “X says it solved Y” can become “Yes”.
* Says “not found” for about 1 in 10 answers that are actually in the results.
* Sometimes quotes a neighbouring fact (another year’s winner). English only.

**Training:** LoRA rank 16, about 9k tool-calling conversations on an RTX 4060 laptop. Search results are Wikipedia and FineWeb pages, every answer and quote checked to appear verbatim, about a quarter teach “not found”. My local qwen 3.8 flash next helped prepare the training data and questions as well!

**Run:**

    llama-server -hf banyohavlusu23/Obsessive-1.7B-GGUF --jinja -c 16384

Temperature 0. Configs for Pi, OpenCode and Hermes plus a keyless Python script are in the repo.

HF: [https://huggingface.co/banyohavlusu23/Obsessive-1.7B-GGUF](https://huggingface.co/banyohavlusu23/Obsessive-1.7B-GGUF)  (Update, the previous link was broken, this is the working one) 

LFM Open License (inherited from the base).

Feedback welcome, especially failure cases.
```



document fetched 2026-11-08 00:34 · canonical https://reddit.com/r/LocalLLaMA/comments/1xq4yju/obsessive17b_a_11_gb_websearch_subagent
