Add wiki_partial indices: BM25, SPLADE++, BGE, Contriever, Qwen3-Embedding-4B
#2
by HeydarS - opened
Adds prebuilt indexes over the partial Wikipedia corpus (wiki_partial.jsonl, 5,910,177 passages) under indices/wiki_partial/, next to e5_partial/. Index files only (no embedding memmaps).
| Folder | Retriever | Settings |
|---|---|---|
bm25_partial/ |
BM25 (Lucene) | positions, doc vectors and raw docs stored |
spladepp_partial/ |
SPLADE++ naver/splade-cocondenser-ensembledistil |
Lucene impact index, max_length 256 |
bge_partial/ |
BAAI/bge-large-en-v1.5 |
FAISS Flat, CLS pooling, max_length 512, d=1024 |
contriever_partial/ |
facebook/contriever-msmarco |
FAISS Flat, mean pooling, max_length 512, d=768 |
qwen3_emb_4b_partial/ |
Qwen/Qwen3-Embedding-4B |
FAISS Flat, last-token pooling, max_length 512, d=2560; 2 parts, cat qwen3_emb_4b_Flat.index.part_* > qwen3_emb_4b_Flat.index |
FAISS row i is line i of wiki_partial.jsonl. The README table and the reassembly instructions are updated.
Compatibility
- The Lucene indexes were built with Pyserini 2.3.0 (Lucene 10 codec,
Lucene104); reading them needs a Pyserini release on Lucene 10. The existingwiki/andecommerce/Lucene indexes useLucene99. - SPLADE++ term weights are quantized as
round(100 * w), as in Pyserini's prebuilt SPLADE indexes; query weights must use the same scale.
Checks (all five indexes): 5,910,177 documents each; 50 random passages used as queries retrieve themselves within the top 10 (rank 1 for 78-90%); 50 wiki2_test queries run end to end through the retrievers.
mahtaa changed pull request status to merged