Skip to content

feat: FineWeb-10B parquet upload and search examples - #175

Open
qdrant-cloud-bot wants to merge 1 commit into
devfrom
feat/fineweb-10b-dataset
Open

feat: FineWeb-10B parquet upload and search examples#175
qdrant-cloud-bot wants to merge 1 commit into
devfrom
feat/fineweb-10b-dataset

Conversation

@qdrant-cloud-bot

Copy link
Copy Markdown
Contributor

Summary

  • Parquet datasets can expose dense/sparse vectors via vector_column / sparse_column (FineWeb stores embeddings alongside metadata in the same shard).
  • Parts templates accept zero-padded {i:04d} so HuggingFace convert/parquet shards resolve correctly.
  • New examples: upload-fineweb-part, upload-fineweb-10b (parts.count defaults to 1 shard), and search-fineweb.

Usage

bfb upload --example upload-fineweb-part -b 64 -p 8 -t 4 --uri http://localhost:6334
bfb search --example search-fineweb -n 10k -p 8 -t 4 --search-limit 10 --uri http://localhost:6334

Raise parts.count in upload-fineweb-10b.yaml to load more of train-part0 (each shard ~5 GB).

Test plan

  • cargo test (182 passed)
  • cargo clippy --tests -- -D warnings
  • Real FineWeb shard smoke read via FINEWEB_SAMPLE (single local file only)
  • Manual upload of one shard against a local Qdrant (optional; ~5 GB download)

Add parquet vector_column/sparse_column support and zero-padded {i:04d}
parts templates so bfb can ingest Qdrant/FineWeb-10B dense+sparse
embeddings from the same shards, with parts.count to limit how much of
the corpus is loaded.
@qdrant-cloud-bot
qdrant-cloud-bot changed the base branch from master to dev September 6, 2026 09:51
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant