--- license: odc-by task_categories: - question-answering - text-generation language: - ar - cs - zh - da - de - el - fa - fr - hu - id - it - ja - nl - pl - pt - ru - es - sv - tr - vi language_bcp47: - arb-Arab - ces-Latn - cmn-Hani - dan-Latn - deu-Latn - ell-Grek - fas-Arab - fra-Latn - hun-Latn - ind-Latn - ita-Latn - jpn-Jpan - nld-Latn - pol-Latn - por-Latn - rus-Cyrl - spa-Latn - swe-Latn - tur-Latn - vie-Latn size_categories: - 100M **Notice:** This repository contains the **unfiltered raw generation** of **FineWeb2-HQ-QA** across all 20 languages. The filtering pipelines and curation code used for experiments will be released separately soon. **FineWeb2-HQ-QA** is a large-scale multilingual synthetic dataset spanning **188.8 million web documents** and their generated question-and-answer pairs. Unlike traditional multilingual synthetic data created by translating English source data, FineWeb2-HQ-QA generates QA pairs directly in the target language grounded in native web documents sampled from [FineWeb2-HQ](https://huggingface.co/datasets/epfml/FineWeb2-HQ). We are releasing the raw, unfiltered data so researchers can experiment with custom filtering heuristics, tailor pipelines to specific language scripts, or refine individual QA pairs directly. --- ## Dataset Summary - **Type:** Raw / Unfiltered Generations - **Input Document Count:** 188,826,048 documents - **Languages:** 20 languages - **Generator Model:** `Qwen/Qwen3-30B-A3B` (prompted with reasoning enabled; reasoning traces removed in final sample concatenation) - **Format:** Parquet files partitioned by language configuration --- ## Raw Document Counts by Language The numbers below reflect the raw input pool provided in this repository before applying the paper's filtering pipeline: | Configuration (`GlotLID`) | Language | Raw Documents | |:---|:---|---:| | `arb_Arab` | Arabic | 3,353,745 | | `ces_Latn` | Czech | 2,452,481 | | `cmn_Hani` | Mandarin Chinese | 26,076,021 | | `dan_Latn` | Danish | 3,021,637 | | `deu_Latn` | German | 20,759,205 | | `ell_Grek` | Greek | 1,619,179 | | `fas_Arab` | Persian | 1,984,552 | | `fra_Latn` | French | 15,973,025 | | `hun_Latn` | Hungarian | 2,166,540 | | `ind_Latn` | Indonesian | 3,378,722 | | `ita_Latn` | Italian | 10,195,425 | | `jpn_Jpan` | Japanese | 15,043,155 | | `nld_Latn` | Dutch | 6,680,086 | | `pol_Latn` | Polish | 6,640,453 | | `por_Latn` | Portuguese | 7,848,376 | | `rus_Cyrl` | Russian | 34,829,012 | | `spa_Latn` | Spanish | 17,346,026 | | `swe_Latn` | Swedish | 3,306,324 | | `tur_Latn` | Turkish | 2,796,347 | | `vie_Latn` | Vietnamese | 3,355,737 | | **Total** | | **188,826,048** | --- ## Important Usage Notes (Raw Data) Because this is the **unfiltered** release, users should be aware that the raw pool contains: 1. **Incomplete Generations:** A small subset (~1.87%) where generation cut off or the `` closing tag was incomplete. 2. **Potential Language Spills:** Cases (~1.24%) where generation defaulted to English or another language instead of the requested target language. 3. **Repetition Artifacts:** Repetitive questions or n-gram loops (most pronounced in Japanese generations). 4. **Benchmark Contamination:** Potential n-gram overlaps with standard multilingual benchmarks (GlobalMMLU, INCLUDE, XNLI, etc.). If you plan to use this dataset directly for pre-training, we recommend applying language identification (e.g., GlotLID) and Gopher-style repetition filters prior to training. Additionally, we recommend running additional filtering for sensitive content, PII, and repetitions before pre-training. See the [FineWeb2 documentation](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2) for discussions on biases, limitations, and societal impact. --- ## Schema ```text root |-- text: string (nullable = true) |-- id: string (nullable = true) |-- metadata: struct (nullable = true) | |-- context_text: string (nullable = true) | |-- file_path: string (nullable = true) | |-- url: string (nullable = true) | |-- raw_llm_output: string (nullable = true) | |-- questions: string (nullable = true) | |-- source_metadata_json: string (nullable = true) ``` - `text`: Primary document text (matches `context_text`). - `id`: Unique document ID. - `metadata.context_text`: Input web document used as context for generation. - `metadata.questions`: Generated QA pairs for the document. - `metadata.raw_llm_output`: Unprocessed model generation string. - `metadata.url`: Source web URL. - `metadata.source_metadata_json`: Original FineWeb2-HQ metadata serialized as JSON. --- ## Quickstart ```python from datasets import load_dataset # Stream a single language ds = load_dataset( "epfml/FineWeb2-HQ-QA", data_files="data/deu_Latn/*.parquet", split="train", streaming=True, ) sample = next(iter(ds)) print(sample["id"]) print(sample["metadata"]["questions"]) ``` --- ## Dataset origin Being a subset of FineWeb2, this data covers websites over the 2013-2024 time period. FineWeb2 is sourced from the internet at large, it is very likely that some personable identifiable information (PII) will be present, even if the FineWeb2 processing has already anonymized email addresses and public IP addresses. If you find your own PII and would like it removed, please fill out the [FineWeb2 PII removal/opt out form](https://forms.gle/VyNT3ZAUPZjPuWp39). CommonCrawl respects robots.txt at crawl time, but if you are a webmaster and find your website in FineWeb2 and would like to have it removed, you may also use the [FineWeb2 PII removal/opt out form](https://forms.gle/VyNT3ZAUPZjPuWp39). --- ## License The synthetic annotations and dataset curation are released under the [Open Data Commons Attribution License (ODC-By v1.0)](https://opendatacommons.org/licenses/by/1-0/). Underlying web text remains subject to the original source web terms and Common Crawl's Terms of Use.