vidore3_financefr_neomme_260m_li

Multi-vector (late-interaction) embeddings of ViDoRe financefr (vidore/financefr), encoded with Hcompany/NeoMME-260M-Retriever-ST-late at revision 023be2a8ab9d797f5aa76f5bf8b5dde78d819659.

Source data: Hugging Face dataset vidore/vidore_v3_finance_fr at revision 1d808daa08032ffecdf62da151a7f7a8fe2bd0c9, 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

file	dtype	shape	contents
documents.npy	float16 (<f2)	[7,232,957, 128]	every document vector, concatenated document by document (1,765.9 MiB)
doclens.npy	int32	[2,384]	vectors per document; cumsum gives offsets
doc_ids.npy	<U4	[2,384]	original document ids
queries.npy	float32 (<f4)	[1,920, 77, 128]	query vectors, zero-padded at the end (72.2 MiB)
query_lens.npy	int32	[1,920]	true vectors per query, before padding
queries_ids.npy	<U4	[1,920]	original query ids
qrels.test.tsv	text	8,808 rows	TREC qrels, qid \t 0 \t docid \t relevance, no header
gt_top1000.tsv	text	1,920,000 rows	exact MaxSim top-1000, see below
gt_top100.tsv	text	192,000 rows	first 100 ranks of gt_top1000.tsv, same format

All positional indices (the gt_top*.tsv files, and the row order of every .npy file) refer to the order of doc_ids.npy and queries_ids.npy. Reordering either file invalidates the ground truth.

Statistics

documents	2,384
document vectors	7,232,957
vectors per document (min / median / mean / max)	2362 / 3138 / 3034.0 / 3266
queries	1,920
vectors per query (min / median / mean / max)	19 / 36 / 37.3 / 77
queries with at least one qrel	1,920
qrels rows	8,808
embedding dimension	128

Encoding

model	Hcompany/NeoMME-260M-Retriever-ST-late
model revision	023be2a8ab9d797f5aa76f5bf8b5dde78d819659
library	sentence-transformers 6.0.1 MultiVectorEncoder (transformers 5.3.0, torch 2.11.0+cu130)
document compute dtype	float16 (model weights loaded at this dtype for the document pass)
document storage dtype	fp16
query compute dtype	float32 (model weights loaded at this dtype for the query pass)
query storage dtype	fp32
normalization	L2, by the model's own Normalize module, before the storage cast
document truncation	none (document_length unset), so the model limit of 16,384 tokens applies; longest document here 3,266 vectors
query truncation	none (query_length unset)
document skiplist	none (empty skiplist_words): every document token is kept, so doclens is the real token count
document input	page image, one vector per image patch plus layout tokens; processor default resizing
query input	query text, stripped of surrounding whitespace, formatted by the model's own query prompt/template
query vectors	every vector the model emits for the query is kept, including any query-expansion tokens its template adds; query_lens counts them all
document padding	none: documents.npy holds real vectors only, sum(doclens) == n_tokens
query padding	rows at or beyond query_lens[i] in queries.npy[i] are exactly zero
token_ids	not provided: image-patch vectors have no vocabulary ids (only placeholder ids)

Ground truth: gt_top1000.tsv and gt_top100.tsv

Exact brute-force MaxSim top-1000 per query over the full corpus, from the vectors in this repo. gt_top100.tsv holds the first 100 ranks per query of the same lists (the original layout of these exports).

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.

Retrieval quality

Sanity check of the vectors, not a leaderboard number: gt_top1000.tsv (exact MaxSim over the full corpus) scored against qrels.test.tsv with ir_measures.

nDCG@10	MRR@10	Success@5	Recall@100	Recall@1000	MAP@1000
0.3770	0.4615	0.6120	0.7623	0.9863	0.3265

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, 77, 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 (7232957, 128)
• ✅ doclens.npy dtype/shape — <i4 (2384,)
• ✅ doc_ids.npy is a string array — <U4 (2384,)
• ✅ queries.npy dtype/shape — <f4 (1920, 77, 128)
• ✅ query_lens.npy dtype/shape — <i4 (1920,)
• ✅ queries_ids.npy is a string array — <U4 (1920,)
• ✅ sum(doclens) == n_tokens — 7232957 vs 7232957
• ✅ no empty documents — min doclen 2362
• ✅ len(doc_ids) == len(doclens) == corpus size — 2384, 2384, 2384
• ✅ doc_ids unique
• ✅ query arrays aligned — 1920, 1920, 1920
• ✅ doc and query dim agree — 128 / 128
• ✅ document vectors unit-norm (100k sample) — norm range [0.9995, 1.0006]
• ✅ query vectors unit-norm — norm range [1.000000, 1.000000]
• ✅ all vectors finite
• ✅ gt_top100.tsv has k rows per query — 192000 rows, k=100
• ✅ gt_top100.tsv rows grouped by qidx with ranks 1..k and descending scores
• ✅ gt_top100.tsv indices in range
• ✅ gt_top1000.tsv has k rows per query — 1920000 rows, k=1000
• ✅ gt_top1000.tsv rows grouped by qidx with ranks 1..k and descending scores
• ✅ gt_top1000.tsv indices in range
• ✅ gt_top100.tsv is the first 100 ranks of gt_top1000.tsv

Provenance

exported	2026-09-24
hardware	cpu
revised	2026-09-29: ground truth extended to top-1000 (gt_top1000.tsv, exact MaxSim over this repo's vectors on Tesla V100S-PCIE-32GB); gt_top100.tsv rewritten as its first 100 ranks: 312 rows differ from the previous file, 4 with a different document at that rank, scores moving by at most 0.000005