ningpy/intent-detection-V2.3

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

ningpy/intent-detection-V2.3 is a 7.6 billion parameter intent classifier and safety router fine-tuned from Qwen2.5-7B-Instruct, designed for health application user messages. It categorizes inputs into five states: PASS, REDFLAG, BLOCK, STEER, or COEXIST, with a context length of 32768 tokens. This model excels at identifying health-related user intents and safety concerns, achieving a Safety F1 score of 92.3% and a multi-label exact match of 88.0%. Its primary use is to route user queries in health apps for appropriate automated or human responses.

Loading preview...

Model Overview

ningpy/intent-detection-V2.3 is a specialized 7.6 billion parameter intent classifier and safety router, fine-tuned from the Qwen2.5-7B-Instruct base model. It is specifically designed for processing user messages within health applications, routing them into one of five distinct states: PASS, REDFLAG, BLOCK, STEER, or COEXIST (which signifies an emergency combined with a personal-dose query).

Key Capabilities and Performance

This model offers robust performance in classifying user intent and identifying safety-critical messages:

  • High Safety Detection: Achieves a Safety F1 score of 92.3% for REDFLAG and BLOCK states, with REDFLAG F1 at 94.1%.
  • Multi-label Classification: Features a new multi-label capability with an 88.0% Multi-label Exact Match and 76.2% Coexist recall.
  • Detailed Output Schema: Provides a fixed JSON output with keys for intent, redflag status, redflag_code, block action, block_code, and extracted entities (drugs, diseases, symptoms) verbatim from the user text.
  • Comprehensive State Routing: Defines 5 clear state combinations (PASS, REDFLAG only, BLOCK only, STEER only, COEXIST) to guide downstream actions.
  • Multilingual Training: Trained on 54,404 samples including English, Chinese, and Bruneian Malay, with 57.1% of samples containing back-filled entities.

Use Cases

This model is ideal for health applications requiring:

  • Automated Triage: Routing user queries to appropriate handlers (e.g., normal LLM response, emergency guidance, refusal of advice).
  • Safety Flagging: Identifying potential emergencies or requests for personal clinical judgment.
  • Entity Extraction: Automatically extracting relevant medical entities from user input for further processing.

Finetunes
1 models