FinHeadlineMix

# FinHeadlineMix **500,000 unique synthetic financial news headlines**, generated with NVIDIA NeMo Data Designer and deduplicated with NeMo Curator. ## Dataset description FinHeadlineMix is a large-scale synthetic dataset of financial news headlines spanning 13 topic categories (12 domain topics plus a catch-all *Other* class). All headlines were generated using NVIDIA NeMo Data Designer, with NVIDIA's [**Nemotron-3-Nano-30B-A3B**](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16) model configured as the generation model. It was produced by an iterative generation–curation pipeline that ran for 82 iterations on a single 8-way NVIDIA B200 node. Each iteration generated a batch of headlines, used NVIDIA NeMo Curator for semantic deduplication, selected maximally diverse few-shot examples via a *farthest-from-centroid* strategy, and corrected the category distribution before the next batch. The dataset is intended as a pre-training or fine-tuning corpus for a range of financial NLP tasks, including headline classification, headline generation, and representation learning. It supports the full spectrum of fine-tuning workflows, from parameter-efficient adaptation with LoRA to model compression and knowledge distillation, where large teacher models are compressed into smaller, production-ready student models. ## Data fields | Column | Type | Description | |--------|------|-------------| | `id` | `string` | UUID v4 — unique identifier for each headline | | `headline` | `string` | Synthetic financial news headline (5–38 words, avg ~15) | ## Generation methodology | Stage | Tool | Details | |-------|------|---------| | Generation | NeMo Data Designer + Nemotron-3-Nano-30B-A3B | 35K–50K headlines per batch; category-weighted sampling with per-category few-shot prompts | | Quality filter | Rule-based | Min 5 words, max 100 words, special-character ratio < 25% | | Deduplication | NeMo Curator | MiniLM-L6-v2 embeddings (384-dim), 500 K-means clusters, 90% cosine-similarity threshold — applied globally against the full accumulated corpus each iteration | | Few-shot selection | Farthest-from-centroid + semantic filter | Top-3 most semantically distinctive headlines per category; candidates with ≥ 80% cosine similarity to any previously used example are rejected | | Distribution correction | Ratio-based weight adjustment | Under-represented categories boosted proportionally for the next batch | ## Usage ```python from datasets import load_dataset ds = load_dataset("nvidia/FinHeadlineMix") print(ds["train"][0]) # {'id': '83cd51a2-...', 'headline': 'Industrial manufacturers lower 2024 profit guidance as supply chain constraints persist'} ``` ## Related resources - [NeMo Data Designer](https://github.com/NVIDIA-NeMo/DataDesigner) — open-source synthetic data generation library - [NeMo Curator](https://github.com/NVIDIA/NeMo-Curator) — open-source scalable data curation and deduplication library - [Nemotron-3-Nano-30B-A3B](https://build.nvidia.com/nvidia/nemotron-3-nano-30b-a3b) — open MoE model used for generation - [Scaling Laws for Task-Specific LLM Distillation](https://arxiv.org/abs/2606.24747) — research paper leveraging this dataset to derive distillation scaling laws - [AI Model Distillation for Financial Data](https://developer.nvidia.com/blog/build-efficient-financial-data-workflows-with-ai-model-distillation/) — companion developer example for downstream distillation ## Citation ```bibtex @dataset{findistill_finheadlinemix_2026, title = {FinHeadlineMix: Large-Scale Synthetic Financial News Headlines}, author = {Desai, Dhruv and Ghita, Lavinia}, year = {2026}, publisher = {NVIDIA}, url = {https://huggingface.co/datasets/nvidia/FinHeadlineMix} } ``` ## License The FinHeadlineMix dataset is licensed under the Creative Commons Attribution 4.0 International License ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/legalcode)).