BharatGPT

Data Curation for English Foundational Model

RedPajamaV2 directory contains scripts for processing and filtering of RedPajamaV2 dataset.

Overall pipeline is as follows:

URL Filtering scripts/utils/url_filtering.py: The first step of filtering is URL based. We use a blocklist of URLs known to contain inappropriate content. This list is taken from blocklistproject. It contains categories for advertisements, gambling, adult content, etc. All documents having URL source domain in these lists is filtered-out from any further processing.
Quality Filtering scripts/utils/quality_signals_filtering.py: Three processing steps take place here:
- Duplicate Removal: Exact duplicates as computed by RPv2 are removed at this step.
- Filtering by quality-signals: Based on quality annotations provided by RPv2, we filter-out documents that do not fit our various criteria. The report contains a detailed description of these.
- Filtering by repetition-removal: We filter documents that contain too much repetition.
Minhash Deduplication scripts/utils/minhash_clusters_filtering.py: Here we remove fuzzy-duplicate documents identified after minhash-clustering.

These steps are directly inspired from RefinedWeb. This has closed-source implementation and authors released only a fraction of the entire dataset.

Data_Info contains a script for fast downloads from HuggingFace. It is taken from this gist. Its usage is in download-notes.md.