PLJIANGG05/qwen35-4b-danger-intent-opendata-qwensynth-79288-v5

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-79288-v5 is a 4.5 billion parameter Qwen3.5-based model specifically fine-tuned for binary classification of user input as 'SAFE' or 'UNSAFE'. This research classifier, developed by PLJIANGG05, leverages a diverse training pool of 79,288 records, including open-source datasets and Qwen-generated synthetic data. It is designed for text-input safety classification, providing a risk score and truncation indicator, with a maximum classification context of 1,024 tokens.

Loading preview...

Model Overview

This model, PLJIANGG05/qwen35-4b-danger-intent-opendata-qwensynth-79288-v5, is a 4.5 billion parameter Qwen3.5-based classifier designed to categorize user input as either SAFE or UNSAFE. It is a specialized research tool, not a general-purpose assistant or certified safety system, focusing exclusively on text-input safety classification.

Key Capabilities and Features

  • Binary Safety Classification: Determines if user input is SAFE or UNSAFE.
  • Extensive Training Data: Fine-tuned on a 79,288-record dataset comprising six public-source datasets (e.g., Aegis V2, OR-Bench, AdvBench) and 29,833 corrected Qwen-generated synthetic samples.
  • Scoring Output: Provides a binary prediction, raw margin, risk score, threshold, and truncation indicator.
  • Context Handling: Processes a maximum classification context of 1,024 tokens, with longer inputs retaining roughly 75% of the beginning and 25% of the end.
  • Multilingual Presence: English and Chinese are present in the training data, though expanded comparisons are currently English-only.

Use Cases and Limitations

This model is primarily intended for research and development in content moderation and safety filtering for text-based inputs. It is crucial to note its limitations:

  • Not a General-Purpose Assistant: It does not perform general language generation or understanding tasks.
  • No Image/Video Safety Claims: Training and evaluation were exclusively on text; no claims are made regarding image or video safety.
  • Research Classifier: Users should validate its performance on independent data and maintain human review for critical decisions.
  • Context Truncation: Inputs exceeding 1,024 tokens may lose safety-relevant context due to truncation.

Finetunes
1 models