| --- |
| license: odc-by |
| language: |
| - en |
| task_categories: |
| - text-classification |
| size_categories: |
| - 100K<n<1M |
| source_datasets: |
| - HuggingFaceFW/fineweb |
| pretty_name: FineWeb-Edu scores via NeMo Curator on HF Jobs |
| tags: |
| - fineweb |
| - nemo-curator |
| - hf-jobs |
| - educational-quality |
| - demonstration |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/*/*.jsonl |
| --- |
| |
| # FineWeb-Edu scores via NeMo Curator on HF Jobs |
|
|
| 400,000 English web documents from [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) (the first 400k rows of the `sample/10BT` subset, file `000_00000.parquet`), scored for educational quality with the [FineWeb-Edu classifier](https://huggingface.co/HuggingFaceFW/fineweb-edu-classifier) running as [NVIDIA NeMo Curator](https://github.com/NVIDIA-NeMo/Curator)'s `FineWebEduClassifier` stage on [Hugging Face Jobs](https://huggingface.co/docs/huggingface_hub/en/guides/jobs). |
|
|
| This is a **demonstration artifact, not a curated corpus**: it exists as the worked example for a [tutorial on running Curator array workloads on HF Jobs](https://github.com/davanstrien/nemo-curator/tree/tutorial/hf-jobs-array/tutorials/hf_jobs). The scores are real and reusable, but the document selection is simply "the first file of a FineWeb sample" — don't read anything into what's included. |
|
|
| ## How it was produced |
|
|
| The 400k rows were split into 16 input files (25k rows each) and processed by **4 independent GPU Jobs (`l4x1`)** using Curator's own Slurm-array sharding machinery — the `NEMO_CURATOR_SLURM_ARRAY_*` environment variables, which turn out to be scheduler-agnostic — with **zero changes to the library**. File-to-shard assignment is hash-based, which is why the four shard directories hold 3/4/5/4 files rather than 4/4/4/4. |
|
|
| One shard was interrupted mid-run (simulating a preemption). Curator's stock retry planner (`tutorials/slurm/retry_array.py`, unmodified) identified the missing shard from completion manifests written to a Hugging Face storage bucket, and a retry wave completed it — the retried shard re-derived exactly its original file assignment. |
|
|
| Throughput was 292–303 rows/s per L4; the whole run, including the interrupted shard and its retry, cost ≈ $0.50. |
|
|
| ## Data structure |
|
|
| One `train` split, 16 JSONL files under `data/shard{0..3}/`. |
|
|
| | column | type | description | |
| |---|---|---| |
| | `id` | string | FineWeb document id (`urn:uuid:...`) | |
| | `text` | string | document text, unchanged from FineWeb | |
| | `fineweb-edu-score-float` | float | raw classifier score (≈0–5) | |
| | `fineweb-edu-score-int` | int | rounded score | |
| | `fineweb-edu-score-label` | string | `low_quality` / `high_quality` | |
|
|
| Score distribution over all 400,000 rows: |
|
|
| | score (int) | label | rows | |
| |---|---|---| |
| | 0 | low_quality | 48,061 | |
| | 1 | low_quality | 248,737 | |
| | 2 | low_quality | 80,309 | |
| | 2 | high_quality | 81 | |
| | 3 | high_quality | 19,832 | |
| | 4 | high_quality | 2,961 | |
| | 5 | high_quality | 19 | |
| |
| The expected FineWeb skew: 94% of documents score ≤2, and only 5.7% carry `high_quality`. |
|
|
| > [!NOTE] |
| > The `label` column is thresholded on the **float** score, not the rounded int — hence the 81 rows with `score-int = 2` but `label = high_quality`. If you filter, use `fineweb-edu-score-float` or the label; mixing them with the int column will give slightly inconsistent subsets. |
| |
| ## Using it |
| |
| ```python |
| from datasets import load_dataset |
|
|
| ds = load_dataset("davanstrien/fineweb-edu-showcase", split="train") |
| high = ds.filter(lambda r: r["fineweb-edu-score-label"] == "high_quality") |
| ``` |
| |
| Plausible uses beyond the tutorial: a ready-made small testbed for score-threshold experiments against the FineWeb-Edu classifier, or a quick source of quality-stratified English web text. For serious educational-quality filtering use [FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) itself — that's the full-scale version of exactly this pipeline. |
| |
| ## Licence and credit |
| |
| Text is from FineWeb and carries its [ODC-By 1.0](https://opendatacommons.org/licenses/by/1-0/) licence (with CommonCrawl's terms of use upstream of that). The score columns are model outputs from the FineWeb-Edu classifier. |
| |
| Source data by [HuggingFaceFW](https://huggingface.co/HuggingFaceFW) (FineWeb, FineWeb-Edu classifier); pipeline stage by [NVIDIA NeMo Curator](https://github.com/NVIDIA-NeMo/Curator). Scored and repackaged by [Daniel van Strien](https://huggingface.co/davanstrien). |
| |
| ```bibtex |
| @inproceedings{penedo2024fineweb, |
| title={The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale}, |
| author={Penedo, Guilherme and Kydl{\'\i}{\v{c}}ek, Hynek and Lozhkov, Anton and Mitchell, Margaret and Raffel, Colin and Von Werra, Leandro and Wolf, Thomas and others}, |
| booktitle={NeurIPS Datasets and Benchmarks}, |
| year={2024} |
| } |
| ``` |
| |
