principled-intelligence/claim-extractor-2B-q-2605

ClaimExtractor-2B-q-2605: Structured Extraction for AI Conversations

ClaimExtractor-2B-q-2605, developed by Principled Intelligence, is a 2.3 billion parameter small language model (SLM) optimized for transforming AI conversations into structured, auditable claims and intents. Unlike larger LLMs, this model prioritizes consistent, low-latency extractions, making it cheaper and faster for production deployments. It processes the last message of a conversation, optionally using an AI service description, to extract self-contained claims (Factoid, Capability, User Assertion, Unverifiable) and user intents.

Key Capabilities

• Decontextualized Extractions: Generates claims and intents that stand alone, simplifying consumption by fact-checkers, intent routers, and audit pipelines.
• Structured Outputs: Categorizes claims into four subtypes and identifies explicit user intents.
• Optimized for Production: Designed for inline, per-turn extraction with a focus on low latency and memory efficiency, suitable for single consumer-grade GPUs.
• Improved Quality with Service Description: Benefits significantly from a structured AIServiceDescription for more precise extractions.

Good For

• Fact-checking pipelines: Verifying Factoid claims against authoritative data.
• Capability auditing: Ensuring assistant Capability claims align with system functionalities.
• Intent routing: Directing user requests to appropriate tools or human agents.
• Compliance & brand monitoring: Flagging Unverifiable claims or banned language.
• Long-term analytics: Analyzing trends in user intents and assistant behavior.