flowxai/semantic-mapper

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The flowxai/semantic-mapper is a 4 billion parameter language model, LoRA-fine-tuned from Qwen/Qwen3-4B by FlowX.AI, designed to extract structured ontology JSON from regulatory and legal text. It processes single chunks of text in English, French, German, and Romanian across banking, insurance, logistics, and labor domains, mapping them to structural, semantic, and governance facets, including a 252-concept controlled vocabulary.

Loading preview...

FlowX Semantic Mapper: Regulatory Text to Ontology

The flowxai/semantic-mapper is a 4 billion parameter model, fine-tuned from Qwen/Qwen3-4B by FlowX.AI, specifically engineered to transform chunks of regulatory or legal text into structured JSON ontologies. It operates across English, French, German, and Romanian, covering banking, insurance, logistics, and labor domains.

Key Capabilities

  • Structured Ontology Extraction: For each text chunk, it generates a JSON output with three facets:
    • Structural: Identifies the clause's location (source ID, document type, hierarchy).
    • Semantic: Provides free-text domain_tags, controlled-vocabulary concepts (from a 252-concept taxonomy), and entity triples (actor, action, object, constraint).
    • Governance: Includes policy references (PDP.<domain>.<rule>) and an escalation_trigger for downstream systems.
  • Controlled Vocabulary: A central design choice is the use of a 252-concept controlled taxonomy for concepts, significantly improving learnability and F1 scores (from 0.24 to 0.54) compared to open free-text concepts.
  • Pipeline Integration: Designed as the first stage in a compliance pipeline, feeding into a policy layer and the flowxai/sentinel-gate escalation model.
  • Multilingual Support: Supports EN, FR, DE, RO, with EN being the strongest.

Good for

  • Automated Regulatory Compliance: Ideal for organizations needing to process large volumes of regulatory text to identify obligations, policies, and potential escalation points.
  • Consistent Tagging: Provides consistent, joinable tags for regulatory clauses, enabling robust downstream analysis and policy enforcement.
  • On-device Deployment: Available in multiple formats including fp16 safetensors, MLX-quantized (int4/int8) for Apple Silicon, and GGUF (Q8_0/Q4_K_M) for CPU/Ollama, facilitating diverse deployment scenarios.

Limitations

  • Not Legal Advice: The model extracts structure and topics; it does not interpret law or determine compliance.
  • Partial Extraction: concepts F1 is ~0.54 and entities accuracy is ~0.53, meaning extraction is a strong signal but not exhaustive.
  • Single Chunk Processing: Optimized for processing one text chunk at a time; batching multiple clauses degrades output quality.

Finetunes
1 models