mustafabasar/ik-belge-asistani-qwen3.5-4b

VISIONConcurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 18, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The mustafabasar/ik-belge-asistani-qwen3.5-4b is a 4.5 billion parameter Qwen3.5 fine-tuned model designed to structure free-form HR notes into clean JSON. It supports both Turkish and English input, generating JSON output in the corresponding language. This model excels at transforming raw HR documentation from disciplinary meetings, ethics reports, internal audits, and performance reviews into a structured format, while being specifically trained to avoid hallucination and prioritize privacy.

Loading preview...

Overview

The mustafabasar/ik-belge-asistani-qwen3.5-4b is a specialized 4.5 billion parameter Qwen3.5 model, fine-tuned to convert unstructured HR notes into structured JSON. It operates bilingually, producing Turkish JSON for Turkish input and English JSON for English input, maintaining the original language of the content. The model is designed to process various HR document types, including disciplinary records, ethics hotline reports, internal audit findings, and performance consistency checks.

Key Capabilities

  • Bilingual Output Matching: Automatically generates JSON in Turkish or English based on the input language.
  • Hallucination-Averse: Trained to only extract and structure information explicitly stated in the input, flagging unverified details rather than asserting them as facts.
  • Privacy-Conscious: Prioritizes job titles/roles over personal names for unverified third parties, enhancing data privacy.
  • Small and CPU-Deployable: With 4.5 billion parameters, it can run efficiently on CPU-only hardware, making it accessible for various deployment scenarios.
  • Structured HR Document Processing: Transforms raw notes from disciplinary meetings, ethics reports, internal audits, and performance reviews into predefined JSON fields.

Limitations

  • Fixed JSON Field Names: JSON field names remain in Turkish (e.g., bulgu, kanit) even when the input and values are in English, as these are treated as structural constants.
  • Context Window: Has a total context window of 4096 tokens, with 400 tokens reserved for output, which may limit the length of raw notes that can be processed.

Training Details

The model was fine-tuned using LoRA via Unsloth on an unsloth/Qwen3.5-4B base model, trained in float32 precision, and is available as a Q4_K_M GGUF for CPU inference.

Finetunes
1 models