README.md

metadata

license: unknown
language:
  - en
  - fr
  - ar
  - hi
  - es
  - zh
tags:
  - dpo
  - preference-learning
  - reasoning
  - chain-of-thought
  - multilingual
  - post-training
size_categories:
  - 100K<n<1M

Reasoning-Augmented Preference Pairs

140,582 preference pairs across 6 languages and 8 domains, each chosen response paired with a step-by-step reasoning trace.

Every row includes a chosen/rejected preference pair and a reasoning trace behind the chosen response. This supports DPO training on the preference signal and reasoning distillation from the traces, together or separately. Rows come from 13 public instruction and reasoning datasets, filtered to HARD and MEDIUM difficulty, with the five non-English languages holding roughly 10% each.

Highlights

• Reasoning traces. Each row pairs chosen/rejected with a reasoning_trace for the chosen response. The trace explains the preferred answer, not the rejected one.
• Difficulty filtered. HARD and MEDIUM rows only. EASY rows are dropped during selection.
• Multilingual. English plus Arabic, French, Hindi, Mandarin, and Spanish, at roughly 10% each for the five non-English languages.
• Source context. 5 of the 13 sources include the underlying context (financial filing, clinical record, article) in the same language as the prompt.
• Domain coverage. 8 buckets: general instruction-following, medical, finance, science, math, code, safety, and other.

Quick start

from datasets import load_dataset

ds = load_dataset("your-org/preference-pairs")  # replace with the published repo id
print(ds["train"][0])

Dataset summary

Rows	140,582
Columns	prompt, context, chosen, rejected, reasoning_trace, language, domain, source_dataset
Languages	English + Arabic, French, Hindi, Mandarin, Spanish (~10% each)
Domains	8 buckets: general instruction-following, medical, finance, science, math, code, safety, other (legal and summarization fold into other)
Difficulty	HARD and MEDIUM only

Columns

Column	Description
prompt	The task prompt, as sourced from the originating dataset. No separate rewrite or enhancement step this round (see What's different from v1).
context	Supporting source document or data referenced by the prompt (financial filing, clinical record, article, etc.), in the same language as prompt. Present for 5 of the 13 sources.
chosen	The preferred response to the prompt.
rejected	The dispreferred response to the same prompt. Pair with chosen for DPO-style preference training.
reasoning_trace	A step-by-step reasoning trace behind the chosen response specifically, not rejected.
language	Detected language of prompt (ISO 639-1 code where available).
domain	Coarse subject-area bucket derived from the originating source dataset.
source_dataset	Which of the 13 public datasets the row came from.

Curation recipe

1. Source aggregation. 13 public datasets spanning general instruction-following, medical calculation, finance QA, science, math, code, legal scenarios, summarization, and safety (see Source composition).
2. Difficulty-based selection. HARD and MEDIUM only. EASY rows are dropped.
3. Reasoning + preference pairs. Each row carries a chosen/rejected pair plus a reasoning_trace behind the chosen response.

Distributions

Language

Domain

Source composition

Source dataset	Domain bucket	Rows kept
WizardLM/WizardLM_evol_instruct_V2_196k	general	52,143
ncbi/MedCalc-Bench-v1.2	medical	23,373
ibm-research/finqa	finance	16,598
nvidia/Nemotron-Science-v1	science	14,899
AI-MO/NuminaMath-CoT	math	8,982
nvidia/Nemotron-SFT-v3-IF	general	6,952
nvidia/Nemotron-Competitive-Programming-v1	code	4,390
qiaojin/PubMedQA	medical	4,130
nvidia/Nemotron-SFT-Safety-v1	safety	3,773
nvidia/Nemotron-Math-v2	math	3,005
ise-uiuc/Magicoder-OSS-Instruct-75K	code	1,380
relai-ai/legal-scenarios-SCOTUS-2024-decisions	legal	491
EdinburghNLP/xsum	summarization	466
Total		140,582

License and attribution

This dataset aggregates 13 public source datasets, each under its own license. Downstream use is subject to the most restrictive applicable source license. Confirm and list per-source terms before publishing.