README.md

metadata

language:
  - es
tags:
  - reasoning
  - venezuela
  - evaluation
  - cultural-context
license: apache-2.0

Venezuelan Reasoning Benchmark

This dataset is designed to evaluate complex reasoning capabilities in Venezuelan Spanish, focusing on tasks that require cultural context and local knowledge.

Dataset Details

Description

This benchmark consists of reasoning tasks that are explicitly related to Venezuelan culture, geography, history, and customs. It is derived from the ScaleAI MultiNRC dataset.

Curation Process

1. Source: Extracted from the original ScaleAI NRC dataset.
2. Filtering: Used Qwen 2.5 Coder 32B to identify and select evaluations that explicitly referenced "Venezuela" or related cultural concepts.
3. Refinement: The original english_reasoning fields were manually translated and adapted into spanish_reasoning using Gemini 3 Pro.

Data Fields

• task_id: Unique identifier for the task.
• i18n_prompt: The prompt/question in Spanish (Venezuelan context).
• i18n_gtfa: The ground truth final answer in Spanish.
• spanish_reasoning: A step-by-step reasoning chain explaining the answer in Spanish.
• category: The category of the reasoning task (e.g., Cultural Reasoning & Traditions, Mathematical & Logical Reasoning).

Model Performance Leaderboard

Model	Score
claude-opus-4-5-20251101	37.10%
claude-sonnet-4-5-20250929	23.30%
deepseek-r1:70b	11.20%
claude-haiku-4-5-20251001	9.50%
gpt-oss:20b	9.50%
ministral-3:14b	7.80%
qwen3:32b	5.20%
deepseek-r1:32b	4.30%