lewisdog/lfm2.5-350m-cogs-ask

Overview

This model, lfm2.5-350m-cogs-ask, is a 350 million parameter student model derived from LiquidAI/LFM2.5-350M, specialized for Cogitarium retrieval-QA. It handles two primary tasks: decomposition (splitting questions into 1-4 retrieval sub-questions) and synthesis (answering from provided wiki notes with [note-id] citations). The model was trained on a real cogs serialization, including frontier-teacher-distilled grounded Q&A over a real vault, ensuring correct performance on deployment distributions without serving hacks.

Key Capabilities & Features

• Question Decomposition: Accurately breaks down complex questions into manageable sub-questions.
• Grounded Answer Synthesis: Generates answers strictly from provided evidence, including precise [note-id] citations and native handling of [[wikilinks]] within note bodies.
• High Grounded Citation Validity: Achieves approximately 89% grounded citation validity on its training vault, demonstrating strong accuracy in attributing information.
• Optimized for Specific Use Cases: Designed for retrieval-augmented generation (RAG) within the Cogitarium ecosystem, requiring no special serving hacks for citation handling.
• Efficient Performance: Recommended for use with Q8_0 quantization (379 MB, 509 tok/s on GB10), with decompose being flawless even at Q4_K_M (229 MB).

When to Use This Model

This model is ideal for applications requiring fast, accurate, and grounded question answering over a structured knowledge base like Cogitarium. It serves as a strong "fast/small tier" option, particularly when the primary goal is to decompose questions and synthesize answers with verifiable citations from provided notes. While it performs well on its trained domain, users should note that a larger Qwen3-1.7B student model exists for higher quality requirements.