aitf-kpm-ugm/Qwen3-4B-KPM3

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 23, 2026Architecture:Transformer Featherless Exclusive Cold

The aitf-kpm-ugm/Qwen3-4B-KPM3 is a 4 billion parameter Qwen3-based causal language model developed by the AITF-KPM-UGM team. It is specifically finetuned for structured issue analysis from news articles, generating JSON-formatted outputs. This model excels at extracting and organizing information such as main issues, background, recent developments, government statements, and public opinions from news text, with a context length of 32768 tokens.

Loading preview...

Model Overview

aitf-kpm-ugm/Qwen3-4B-KPM3 is a 4 billion parameter language model, finetuned from aitf-kpm-ugm/Qwen3-4B-CPT-Base by the AITF-KPM-UGM team. Its primary function is to perform structured issue analysis on news articles, outputting the analysis in a valid and parseable JSON format. This model is designed to identify and categorize key information from news text, making it highly specialized for public issue analysis.

Key Capabilities

  • Structured JSON Output: Generates analysis in a predefined JSON structure, including fields for main issue, background, recent developments, official government statements, and public figures' opinions.
  • News Article Analysis: Optimized for processing and extracting relevant information from news content.
  • Contextual Understanding: Leverages a base model with a 32768 token context length, allowing for comprehensive analysis of longer news articles.
  • Formal and Objective Tone: Adheres to a formal, objective language, strictly basing analysis on provided news information without external assumptions.

Training and Methodology

The model was finetuned using Supervised Finetuning (SFT) with Low-Rank Adaptation (LoRA) on a synthetic dataset of approximately 9,670 training examples. This dataset, generated using Claude Haiku, was specifically formatted for chat templates, ensuring the model learns to produce structured responses. Training utilized trl library's SFTTrainer with a focus on completion-only loss, meaning the model was optimized to generate accurate assistant responses while ignoring the input prompts during loss calculation.

When to Use This Model

This model is ideal for applications requiring automated, structured analysis of news articles, particularly for public policy monitoring, media analysis, or research that benefits from categorized information extraction. Its ability to output clean JSON makes it suitable for integration into data pipelines and analytical dashboards.