RAS1981/Qwen3-4B-outreach-stage4

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Nov 27, 2025License:apache-2.0Architecture:Transformer Open Weights Warm

RAS1981/Qwen3-4B-outreach-stage4 is a 4 billion parameter Qwen3-based language model, specifically trained as a Russian real-estate qualification agent. This model is the final stage of a 5-stage curriculum, designed to fully internalize complex system prompts, enabling zero-prompt inference. It excels at maintaining multi-turn conversation state, handling fragmented queries, and directing clients towards booking real-estate appointments with fast Time-To-First-Token (TTFT).

Loading preview...

RAS1981/Qwen3-4B-outreach-stage4: Russian Real-Estate Agent

This model is a compact, 4-billion parameter Qwen3-based language model, meticulously trained through a 5-stage curriculum to function as a specialized Russian real-estate qualification agent. Its primary innovation lies in its ability to fully internalize long and complex system prompts, culminating in a Stage-4 version that requires zero prompt scaffolding during inference.

Key Capabilities

  • Autonomous Operation: Acts as a self-contained agent, handling conversations without external system prompts.
  • Real-Estate Qualification: Greets users, collects essential parameters (district, budget, timelines), and qualifies potential clients for new property acquisitions.
  • Robust Conversation Handling: Manages noisy/fragmented queries, objections, and corrections across multi-turn interactions.
  • Goal-Oriented Engagement: Guides clients towards booking calls, meetings, or viewings, adhering to business rules.
  • Fast TTFT & Stability: Delivers rapid Time-To-First-Token and stable multi-turn reasoning, maintaining a consistent sales-oriented tone.
  • Structured Output: Produces concise, operator-style messages, often using bullet points or quick summaries.

Training Methodology

The model underwent a progressive prompt-internalization pipeline:

  • Stage 0 (Domain CPT): Built robust Russian real-estate vocabulary and document-level reasoning.
  • Stage 1-3 (SFT Stages): Progressively internalized system templates, core rules, CTAs, and objection handling, moving from full prompts to highly compressed summaries.
  • Stage 4 (Zero Prompt): Final distillation, achieving full internalization of scripts, tone, policy, and conversation flow, enabling query-only inference.

Recommended Use Cases

This model is ideal for applications requiring an efficient, autonomous, and specialized AI agent for:

  • Automated Russian Real-Estate Lead Qualification: Streamlining the initial client interaction process.
  • Customer Service Automation: Handling common inquiries and guiding users in the Russian real-estate sector.
  • Interactive Sales Support: Providing consistent, rule-based responses to potential buyers.

Finetunes
2 models