RAS1981/qwen3-0.6b-turn-detection-v1

Model Overview

This model, RAS1981/qwen3-0.6b-turn-detection-v1, is a highly specialized 0.8 billion parameter Qwen3-based model developed by RAS1981. Its primary function is to act as a conversational boundary detector for Russian real-estate dialogues, predicting the probability that a user's turn has concluded (<|im_end|>) or is continuing.

Key Features & Methodology

• Base Model: Built upon unsloth/Qwen3-0.6B, ensuring efficiency and strong Russian language support.
• Probability-Based Detection: Unlike traditional binary classifiers, it leverages the model's intrinsic next-token prediction to assign a probability to the End-of-Sequence (EOS) token.
• Performance: Demonstrates high confidence in predicting both complete turns (e.g., >90% probability for <|im_end|>) and incomplete turns (near-zero probability for <|im_end|>).
• Latency: Offers extremely fast inference due to its compact 0.6B parameter size.
• Training: Fine-tuned using Single-Token Loss Masking on a balanced dataset of approximately 20,000 complete and incomplete conversational turns, specifically within the Russian real-estate domain.

Use Cases

This model is ideal for applications requiring precise detection of conversational turn boundaries in Russian, particularly in real-estate contexts. It can be used to:

• Determine when a user has finished speaking to trigger a response from a chatbot.
• Improve dialogue management systems by accurately segmenting user utterances.
• Enhance transcription services by marking natural pauses and turn changes.