PLJIANGG05/qwen35-4b-danger-intent-opendata-qwensynth-supplement-83787-v7

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 31, 2026License:cc-by-nc-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

PLJIANGG05/qwen35-4b-danger-intent-opendata-qwensynth-supplement-83787-v7 is a 4.5 billion parameter Qwen3.5-based model developed by PLJIANGG05, specifically fine-tuned for binary classification of user input as 'SAFE' or 'UNSAFE'. This research classifier is trained on a comprehensive 83,787-record lineage, combining public datasets, Qwen-generated samples, and AI-generated supplements. It excels at text-input safety classification, providing a specialized tool for content moderation and risk assessment.

Loading preview...

Overview

This model, PLJIANGG05/qwen35-4b-danger-intent-opendata-qwensynth-supplement-83787-v7, is a 4.5 billion parameter Qwen3.5-based classifier developed by PLJIANGG05. It is specifically designed for binary classification, determining whether a user input is SAFE or UNSAFE. It is a research classifier, not a general-purpose assistant or a certified safety system.

Key Capabilities

  • Binary Safety Classification: Specializes in classifying text inputs into SAFE or UNSAFE categories.
  • Extensive Training Data: Fine-tuned on an 83,787-record lineage, comprising 79,288 first-stage records and 4,499 supplemental records. This dataset includes six public-source datasets (Aegis V2, OR-Bench, Chinese Do-Not-Answer, AdvBench, HarmBench, JBB-Behaviors), corrected Qwen-generated synthetic data, and AI-generated supplements.
  • Configurable Inference: Supports BF16 full model inference, 4-bit quantization, and CPU inference, with a fixed template for binary classification.
  • Multilingual Presence: Training data includes English and Chinese, though expanded comparisons currently focus on English performance.
  • Provenance and Verification: Includes merge_verification.json and SHA256SUMS.json for integrity checks, and retains the original Qwen3.5-4B base model's architecture for adapter compatibility.

Good For

  • Content Moderation: Identifying potentially unsafe user inputs in applications.
  • Risk Assessment: Providing a programmatic layer for initial safety screening of text.
  • Research in Safety Classification: Serving as a specialized tool for studying and implementing safety filters in LLM applications.

Limitations

  • Context Length: Maximum classification context is 1,024 tokens; longer inputs are truncated, potentially discarding safety-relevant information.
  • Text-Only: Only text-input safety classification was trained and evaluated; no image/video safety claims are made.
  • Not a Guarantee: SAFE classification does not guarantee harmlessness of an input or downstream response. Users must validate on independent data and retain human review for critical decisions.

Finetunes
1 models