MagicCaster/crimeradar-event-merge-qwen3-4b-reasoning-20260629
MagicCaster/crimeradar-event-merge-qwen3-4b is a 4 billion parameter Qwen3-4B-Base model, fully fine-tuned for event merging in the CrimeRadar system. It is specifically designed to identify and group candidate events (ASR-derived dispatch/incident reports) that refer to the same real-world incident. This model excels at reasoning tasks, achieving an F1 score of 0.768 on a 60-item consensus benchmark, matching its teacher models and significantly outperforming prior 8B SFT models.
Loading preview...
Model Overview
MagicCaster/crimeradar-event-merge-qwen3-4b is a 4 billion parameter model based on the Qwen3-4B-Base architecture. It has been fully fine-tuned, without LoRA, to perform event merging for the CrimeRadar system. Its primary function is to analyze a set of candidate events, derived from ASR dispatch/incident reports, and determine which ones correspond to the same real-world incident, subsequently emitting a merged grouping.
Key Capabilities
- Event Merging: Identifies and groups related incident reports, crucial for crime analysis and dispatch systems.
- Reasoning Model: Produces a
<think>...</think>block for internal reasoning before generating a JSON output containingevents_csvandgroups. - High Performance: Achieves an F1 score of 0.768 ± 0.042 on a 60-item consensus benchmark, demonstrating performance comparable to its teacher models (GLM-5.2 + minimax) and significantly surpassing prior Qwen3-8B v2 SFT (0.630 F1) and Qwen3-4B v1 SFT (0.571 F1) models.
- Optimized for Large Inputs: Trained on large-scene-enriched data, with approximately 32% of training rows containing 16 or more events.
Recommended Usage
This model is ideal for applications requiring robust event correlation and incident merging, particularly in domains like public safety or news aggregation. Inference is recommended using vLLM with specific sampling parameters (temperature=0.8, top_p=0.95, top_k=40, presence_penalty=0.5) and a per-request thinking_token_budget to manage potential runaway reasoning loops in complex scenarios.