---
license: cc-by-4.0
task_categories:
- text-generation
- question-answering
- text-classification
language:
- en
- ko
- id
- bn
configs:
- config_name: error-level
  data_files: data/JuICE_error_group_level.csv 
- config_name: response-level
  data_files: data/JuICE_response_level.csv
- config_name: sentence-level
  data_files: data/JuICE_sentence_level.csv
---

# JuICE

## Sources
- **Repository:** https://anonymous.4open.science/r/JuICE
- **HuggingFace:** [juice-cultural-eval/JuiCE](https://huggingface.co/datasets/juice-cultural-eval/JuICE)

## About
We present **JuICE** (Benchmark for LLM-**Ju**dge in **I**dentifying **C**ultural **E**rrors), a multilingual dataset of 7,470 span-level annotations of cultural and linguistic errors, collected from native speakers in long-form LLM responses. It covers 1,050 query-response pairs from four countries (the United States, South Korea, Indonesia, and Bangladesh), in both English and their countries' main languages.

## Data Description
- `qa_id` (key): id of query-response pair
- `country`: country
- `language_type`: `english` or `native`
- `language`: language code (`en`, `ko`, `id`, `bn`)
- `query_text`
- `response_text`
- `error_groups`
    - `error_group_id` (key)
    - `paragraph_idx_list`
    - `thick_categories`
    - `errors`
        - `error_id` (key)
        - `paragraph_idx_list`: index of paragraph in `response_text` where `span` is contained   
        - `detector_id`
        - `detector_type`: `human` or `model`
        - `span`
        - `explanation`
        - `raw_annotation`
        - `thick_category`