jed351/Nemotron-3-Nano-RL-Training-Blend-STEM-Yue-Translated

Nemotron-3-Nano-RL-Training-Blend

This dataset is a translated, parallel version of the nvidia/Nemotron-3-Nano-RL-Training-Blend dataset, excluding the instruction following subset.

Dataset Origin & Sampling

The original NVIDIA dataset blends multiple sources. For this parallel corpus, we sampled 1/3 of the datasets except for the instruction following subset where we extracted all unique prompts from the unsampled dataset to avoid duplicates:

Languages

For each subset, files are provided in three parallel directories:

• —en/: English (Source / Cleaned)
• —yue/: Cantonese (Translated)
• —zh-hk/: Traditional Chinese (Hong Kong) (Translated)

The Law Filter (STEM MCQA Cleaning)

The original nvidia/Nemotron-RL-knowledge-mcqa dataset included multiple-choice questions from various general benchmarks. A significant number of these questions turned out to be related to US Law (e.g., Constitutional Law, Contracts, Civil Procedure, and Torts) rather than pure STEM knowledge.

To clean the dataset:

• —We applied a keyword-based law filter (checking for terms like 刑事, 法律, 法例, 憲法, 州, 合約, 法院, 聯邦, 被告, etc.) on the translations.
• —Any entry that triggered this filter was dropped from the Cantonese and Traditional Chinese translations.
• —To keep the three languages aligned, the corresponding English records for these law questions were also excluded.
• —This process successfully filtered out 2,333 law questions, leaving exactly 2,639 clean STEM multiple-choice questions in the final parallel release of stem_mcqa.

nvidia/Nemotron-RL-agent-workplace_assistant Processing

The original nvidia/Nemotron-RL-agent-workplace_assistant dataset features office-style agent tasks, such as retrieving emails, scheduling meetings, and managing calendars with tool-calling capabilities.

To optimize this subset:

• —Deduplication: The raw dataset contained many duplicate user queries. We deduplicated the subset based on the user prompt, retaining only the 990 unique prompts (and 1 system prompt) in the final parallel release.

nvidia/Nemotron-RL-coding-competitive_coding Processing

The nvidia/Nemotron-RL-coding-competitive_coding subset consists of competitive programming questions.

To process and translate this subset:

• —Deduplication: We extracted the core coding questions and deduplicated them down to 4,291 unique questions (out of 4,812 total records) to optimize translation.
• —Cantonese Translation: The Cantonese translations were generated using the in-development model jed351/cantonese_llm_v2_chat_gsm8k_GRPO_Qwen3-30B-A3B-Think-2507.
• —Traditional Chinese Translation: The Cantonese translations of the core questions were further translated into Traditional Chinese (zh-hk) using the google/gemma-4-26B-A4B-it model on vLLM.
• —Rule-Based Quality Filtering: Competitive-programming statements contain structured data (large numbers, LaTeX math, variable names) that must survive translation unchanged. We ran an automated, rule-based quality check on every translation and dropped any record that failed a hard check — e.g. token-limit truncation, dropped large numbers, unbalanced LaTeX, non-Chinese output, or the model solving the problem in code instead of translating it. Filtering is per stage; because zh-hk is translated from yue, questions dropped at the yue stage are also dropped from zh-hk. English (en) is the clean source and is retained in full. This removed 563 unique `yue` questions and 636 unique `zh-hk` questions, leaving final parallel counts of 4,812 `en` / 4,183 `yue` / 4,098 `zh-hk`.