Obsessive-1.7B: a 1.1GB fine-tuned search subagent that hits 94% recall with zero hallucinations
MuzafferMahi · reddit · 2026-11-07
- A Reddit developer fine-tuned LFM3-1.7B into Obsessive-1.7B (1.1 GB), a subagent that does one job: take a factual question, call websearch once, and return a short answer with a verbatim quote and URL — or say "not found" when the answer isn't there.
Fixed-results benchmark (131 held-out SimpleQA questions, 29 with answers removed):
- Finds answer when present: 50% → 94%
- Says "not found" when absent: 0% → 100%
- Invented answers: 10% → 0%
- Median time per question (RTX 4060 laptop): 5.1s → 2.0s
Live Wikipedia search: reply rate 38% → 100%, multi-fact requests 1/40 → 40/40. The base model edges out on raw correct answers (37/120 vs 24/120) by searching repeatedly, but returns nothing 62% of the time.
Known issues: single-turn only, treats reported claims as facts, misses 1 in 10 present answers, sometimes quotes a neighboring fact, English only.
Training: LoRA rank 16, 9k tool-calling conversations, done on an RTX 4060 laptop using Wikipedia/FineWeb pages with every answer verified verbatim; a quarter of examples teach "not found". GGUF release with configs for Pi, OpenCode and Hermes plus a keyless Python script.
More from coding & agent
- Claude Code Projects collaborative view resembles ChatGPT Spaces; Perplexity next — testingcatalog · 2026-11-07
- Anthropic reportedly building collaborative mode for Claude Code Projects — testingcatalog · 2026-11-07
- Context management is still early: correctness, latency and cost trade off — blaizedsouza · 2026-11-07
- odai framework adds governance and evidence checks to stop agents faking "done" — blaizedsouza · 2026-11-07
- BUI decouples Blender's GPU UI kernel into a standalone runtime: 18MB memory, 12ms cold start — Obijuan_cube · 2026-11-07
- Parcha on Agent Sandboxes: Lessons from Building Agents Since 2023 — blaizedsouza · 2026-11-07