Datasets:

mzoelfakar
/

Mini-GSM8K-Multilingual-Alignment

language:
  - en
  - ar
  - fr
  - es
  - de
  - it
  - pt
  - zh
  - ru
license: apache-2.0
annotations_creators:
  - machine-generated
language_creators:
  - machine-generated
multilinguality:
  - multilingual
source_datasets:
  - openai/gsm8k
task_categories:
  - text-generation
tags:
  - dpo
  - math
  - multilingual
  - alignment
  - gsm8k
  - synthetic
  - translation
  - gpt-oss
  - gpt-oss-20b
  - llama-cpp
dataset_info:
  features:
    - name: prompt
      dtype: string
    - name: chosen
      dtype: string
    - name: rejected
      dtype: string
    - name: language
      dtype: string
    - name: prompt_en
      dtype: string
    - name: chosen_en
      dtype: string
  splits:
    - name: train
      num_bytes: 1488448
      num_examples: 900
  download_size: 747740
  dataset_size: 1488448
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Mini-GSM8K-Multilingual-Alignment

A compact, multilingual DPO (Direct Preference Optimization) alignment dataset designed for training small language models to produce high-quality, step-by-step mathematical reasoning across multiple languages.

Purpose

This dataset was created to improve multilingual math reasoning in small language models such as Al-Khwarizmi-3B — a 3B-parameter math tutor model named after the 9th-century mathematician. The goal is to teach the model to prefer detailed, friendly, well-structured solutions over vague, uninformative responses, across 9 languages.

How It Was Created

Source: 900 samples were drawn from OpenAI/GSM8K (socratic split), filtered to include only samples whose answers contain digits.
DPO Triplet Construction (English):
- Prompt: The math problem
- Chosen: The full socratic step-by-step solution
- Rejected: The solution stripped of all English letters — leaving only digits, symbols, and whitespace — creating an abstract, vague, uninformative response
Multilingual Translation: gpt-oss-20b (F16 GGUF from unsloth/gpt-oss-20b-GGUF, served with llama.cpp llama-server and parallel slots; temperature 0.7, reasoning effort medium) was used to translate the prompt and chosen columns into 8 languages (100 samples each), with careful instructions to:
- Preserve all mathematical symbols ($, <<>>, ####, etc.)
- Add a brief, varying, friendly opening sentence to each solution
- Use context-aware translation (the model receives the original English problem when translating solutions)
English Samples: The final 100 samples remain in English, with friendly opening sentences added to the solutions by the same model.

Creation code is available here

Language Distribution

Language	Code	Samples
Arabic	ar	100
French	fr	100
Spanish	es	100
German	de	100
Italian	it	100
Portuguese	pt	100
Chinese	zh	100
Russian	ru	100
English	en	100
Total		900

Schema

Column	Type	Description
`prompt`	string	Math problem (translated to target language)
`chosen`	string	Full step-by-step solution with friendly opener (translated)
`rejected`	string	Digits-only stripped version of the original English solution
`language`	string	ISO 639-1 language code
`prompt_en`	string	Original English GSM8K problem that `prompt` was translated from
`chosen_en`	string	Original English GSM8K solution that `chosen` was translated from (without the friendly opener)

Usage

from datasets import load_dataset

dataset = load_dataset("mzoelfakar/Mini-GSM8K-Multilingual-Alignment", split="train")

# Filter by language
arabic_samples = dataset.filter(lambda x: x["language"] == "ar")

Creator

Created by Mohamed Zoelfakar as part of the Al-Khwarizmi-3B project for advancing multilingual mathematical reasoning in small language models.