robro612/vidore3_physics_neomme_260m_li

vidore3physicsneomme260mli

Multi-vector (late-interaction) embeddings of ViDoRe physics (vidore/physics), encoded with [Hcompany/NeoMME-260M-Retriever-ST-late](https://huggingface.co/Hcompany/NeoMME-260M-Retriever-ST-late) at revision 023be2a8ab9d797f5aa76f5bf8b5dde78d819659.

Source data: Hugging Face dataset `vidore/vidore_v3_physics` at revision a0de276f515acc044b72cae8de53a44bb5a8f1f5, configs corpus / queries / qrels, split test, loaded with datasets. Document, query and qrel ids are the source's own ids, unchanged.

Every document is one variable-length set of 128-d vectors; every query is one variable-length set of 128-d vectors. Documents and queries are stored at different precisions (fp16 and fp32 respectively), see Encoding.

Files

All positional indices (gt_top100.tsv, and the row order of every .npy file) refer to the order of doc_ids.npy and queries_ids.npy. Reordering either file invalidates gt_top100.tsv.

Statistics

Encoding

Ground truth: gt_top100.tsv

Exact brute-force MaxSim top-100 per query over the full corpus, from the vectors in this repo. No header; tab-separated qidx docidx rank score:

• —qidx: 0-based row into queries_ids.npy / queries.npy
• —docidx: 0-based position into doc_ids.npy / doclens.npy
• —rank: 1-based, descending score
• —score: sum over the query's query_lens[qidx] vectors of max over the document's vectors of the dot product, computed in fp32 with the fp16 document vectors upcast to fp32. Expansion vectors are included in the sum. Printed to 6 decimals.

Self-matches are included. 94 of 1,812 queries are themselves corpus documents with the same id and retrieve that document (typically at rank 1). This is the raw nearest-neighbour list; BEIR's evaluation drops such pairs (ignore_identical_ids), so exclude them before scoring against qrels.

Retrieval quality

Sanity check of the vectors, not a leaderboard number: exact MaxSim over the full corpus scored against qrels.test.tsv with ir_measures, with query-id == doc-id pairs dropped as BEIR does.

Loading

import numpy as np

documents = np.load("documents.npy", mmap_mode="r")      # [n_tokens, 128] float16
doclens = np.load("doclens.npy")                         # [n_docs] int32
offsets = np.concatenate([[0], np.cumsum(doclens)])
doc_ids = np.load("doc_ids.npy")                         # [n_docs] str

def document(i):
    return documents[offsets[i]:offsets[i + 1]]           # [doclens[i], 128]

queries = np.load("queries.npy")                         # [n_queries, 72, 128] float32
query_lens = np.load("query_lens.npy")                   # [n_queries] int32
query_ids = np.load("queries_ids.npy")                   # [n_queries] str

def query(j):
    return queries[j, :query_lens[j]]                     # [query_lens[j], 128]

def maxsim(q, d):
    return (q @ d.astype(np.float32).T).max(axis=1).sum()

Validation

Checks run by the exporter on the files exactly as written here:

• —✅ file set — missing=[] extra=[]
• —✅ documents.npy dtype/shape — <f2 (3920508, 128)
• —✅ doclens.npy dtype/shape — <i4 (1674,)
• —✅ doc_ids.npy is a string array — <U4 (1674,)
• —✅ queries.npy dtype/shape — <f4 (1812, 72, 128)
• —✅ query_lens.npy dtype/shape — <i4 (1812,)
• —✅ queries_ids.npy is a string array — <U4 (1812,)
• —✅ sum(doclens) == n_tokens — 3920508 vs 3920508
• —✅ no empty documents — min doclen 2342
• —✅ len(doc_ids) == len(doclens) == corpus size — 1674, 1674, 1674
• —✅ doc_ids unique
• —✅ query arrays aligned — 1812, 1812, 1812
• —✅ doc and query dim agree — 128 / 128
• —✅ document vectors unit-norm (100k sample) — norm range [0.9994, 1.0006]
• —✅ query vectors unit-norm — norm range [1.000000, 1.000000]
• —✅ all vectors finite
• —✅ gt_top100.tsv has k rows per query — 181200 rows, k=100
• —✅ gt rows grouped by qidx with ranks 1..k and descending scores
• —✅ gt indices in range

Provenance