CodeStrux-Tech/tac-1
CodeStrux-Tech/tac-1 is a specialized language model, fine-tuned from Qwen/Qwen3-4B-Instruct-2507, designed to parse Costa Rican reference-point addresses into structured JSON records. This model excels at extracting anchor names, types, directions, and distances from complex narrative addresses. It achieves high F1 scores on anchor name, direction, and distance parsing, outperforming GPT-4.1 on anchor name F1 in specific evaluations. Its primary use case is to provide structured address data for subsequent geographic grounding by companion systems.
Loading preview...
Overview
CodeStrux-Tech/tac-1 is a specialized model for parsing Costa Rican reference-point addresses. Unlike traditional street addresses, Costa Rican addresses are narrative descriptions (e.g., "100 meters south of the old fig tree in San Pedro"). This model transforms these narratives into a structured, leg-based JSON format, extracting components like anchor name, anchor type, direction, and distance for each leg.
Key Capabilities
- Specialized Address Parsing: Converts complex Costa Rican narrative addresses into structured JSON.
- High Accuracy: Achieves a parse rate of 1.000 on evaluation sets and high F1 scores for
anchor_name,direction, anddistance. - Outperforms GPT-4.1: Demonstrates superior
anchor_nameF1 scores compared to GPT-4.1 on thegold-47evaluation set. - Zero Geographic Overfit: Maintains 1.000 F1 scores on never-trained districts, indicating robust generalization.
- Hybrid System Component: Designed to work in conjunction with a separate grounding package (
tico) for resolving addresses to geographic coordinates.
Good For
- Applications requiring Costa Rican address processing: Ideal for systems needing to understand and process the unique address format of Costa Rica.
- Structured data extraction from natural language: Specifically for reference-point address narratives.
- Integration into geographic information systems: Provides the initial parsing layer for subsequent geocoding and routing tasks.