README.md

metadata

pretty_name: CultureConverse-DS
tags:
  - culture
  - dialogue
  - multilingual
  - synthetic
configs:
  - config_name: global_english
    default: true
    data_files:
      - split: train
        path: data/global_english/train-*.jsonl.gz
      - split: test
        path: data/global_english/test-*.jsonl.gz
  - config_name: native
    data_files:
      - split: train
        path: data/native/train-*.jsonl.gz
      - split: test
        path: data/native/test-*.jsonl.gz

CultureConverse-DS

CultureConverse-DS contains scored, multi-turn assistant dialogue episodes for culturally grounded assistance in East and Southeast Asia.

Contents

• Total records in this export: 288905.
• Configs: global_english and native.
• Splits: train contains oracle-guided gold-mode dialogues; test contains blind evaluation-mode benchmark episodes.
• Format: one stripped CultureConverse evaluation record per JSONL row.
• manifest.json records source paths, counts, generated shard files, and stripped fields.

Loading

from datasets import load_dataset

en = load_dataset("Incomple/CultureConverse-DS", "global_english")
native = load_dataset("Incomple/CultureConverse-DS", "native")

Source Mapping

Config	Split	Source key	Source mode	Records
global_english	train	train_gold_en	gold	137144
native	train	train_gold_native	gold	137151
global_english	test	test_eval_en	eval	7305
native	test	test_eval_native	eval	7305