Nemotron-RL-QA-Abstention-v1

**Nemotron-RL-QA-Abstention-v1** - License: cc-by-4.0 - Language: en - Task Categories: reinforcement-learning, question-answering, text-generation - Tags: abstention, question-answering, hotpotqa, software-engineering, health, law, rl, rlvr - Configs: default train split at data/train.jsonl - Domain: multi-domain question answering, abstention - Modality: text - Capability Breakdown: Abstention-aware factoid question answering [100%] - Source: Hybrid: Automated, Manually Collected, Synthetic - Size Bin: <10K - Associated Model Release: Nemotron Ultra ## Dataset Description: Nemotron-RL-QA-Abstention-v1 is a multi-domain reinforcement learning with verifiable rewards (RLVR) dataset for training LLMs to abstain when uncertain. It is intended for an abstention-aware RLVR environment where rewards measure answer correctness and appropriate abstention behavior. The model is expected to answer factoid questions when it knows the answer, and to output `\\boxed{[IDK]}` when it does not know the answer or needs more context/tools, rather than hallucinating. The dataset combines HotpotQA, Go standard library documentation, Wikipedia medical articles, and Wikipedia legal articles. HotpotQA is included as a component of this dataset: the training file contains 450 HotpotQA examples, along with 900 Go documentation examples, 900 Health Wikipedia examples, and 900 Law Wikipedia examples. This release excludes CRAG Finance. CRAG Finance was present in an earlier five-source variant but is not included in this dataset. This dataset is ready for commercial or non-commercial uses. ## Dataset Owner(s): NVIDIA Corporation ## Dataset Creation Date: Created on: 04/09/2026 Last Modified on: 04/09/2026 ## Version: Nemotron-RL-QA-Abstention-v1 <br> ## License/Terms of Use: This dataset is licensed under [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/). ## Intended Usage: This dataset is intended for: * Reinforcement learning with verifiable rewards (RLVR) of LLMs on abstention-aware factoid question answering. * Training models to answer in `\\boxed{answer}` format when confident. * Training models to output `\\boxed{[IDK]}` instead of guessing when uncertain. * Multi-domain QA training across HotpotQA, software engineering, health, and law. * Reducing hallucination by rewarding abstention when a question cannot be answered reliably. ## Dataset Characterization ### Dataset Composition and Generation #### Problem Sources The dataset contains: * HotpotQA public benchmark examples. * Go standard library documentation examples. * Health QA examples from Wikipedia medical articles. * Law QA examples from Wikipedia legal articles. CRAG Finance is not included in this release. #### Curation and Filtering Source documents are collected and long documents are chunked into approximately 800-token passages. Each chunk is passed to Nemotron-3-Super-120B to generate six grounded factoid QA pairs. Generated pairs are filtered by answer length, question format, and `\\boxed{answer}` validity. Questions are deduplicated against the AA-Omniscience evaluation set using 3-gram Jaccard similarity with threshold 0.3. The training file contains no assistant responses; records are system/user inputs for online RL rollouts. Each record preserves the source QA fields, including the target `answer`, and includes normalized release fields for UUID, license, tools, and messages. #### Dataset Fields The JSONL file contains the following top-level fields: * `id`: Source record identifier. * `question`: User-facing question. * `answer`: Target answer for comparison and reward evaluation. * `agent_ref`: Gym agent reference metadata. * `responses_create_params`: Original Responses API parameters, including the system/user input. * `messages`: Two-message conversation with a system instruction and user question. * `tools`: Tool definitions for the record; empty for this dataset. * `license`: Release license metadata. * `uuid`: Normalized record UUID, matching `id`. * `metadata`: Metadata object containing `uuid`. * `domain` and `source`: Source-domain metadata when present. The system message instructs the model to answer with only the answer in `\\boxed{answer}` format, or to output `\\boxed{[IDK]}` if it does not know the answer or needs more context/tools. **Data Collection Method**<br> * Hybrid: Automated, Manually Collected, Synthetic <br> **Labeling Method**<br> * Hybrid: Manually-Labelled, Synthetic, Automated. The dataset uses public benchmark answers, synthetic QA generation with Nemotron-3-Super-120B, rule-based filtering, and deduplication. <br> ## Dataset Format Language: English (en) Modality: Text Format: JSONL Structure: Gym-format records with source ID, question, target answer, system prompt, user question, and normalized release metadata. No assistant response is included; rollouts are generated online during RL training. ## Dataset Quantification | Component | Samples | |-----------|---------| | HotpotQA | 450 | | Go documentation QA | 900 | | Health Wikipedia QA | 900 | | Law Wikipedia QA | 900 | | Total | 3,150 | Total Disk Size: 1.7MB ## Reference(s): * [HotpotQA](https://HotpotQA.github.io) * [Go documentation](https://go.dev/doc/) * [Wikipedia](https://www.wikipedia.org/) ## Ethical Considerations: NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse. Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).