README.md

---
pretty_name: nfcorpus_answerai_colbert_small
license: cc-by-sa-4.0
tags:
- multi-vector
- late-interaction
- colbert
- embeddings
- retrieval
- text
---

# nfcorpus_answerai_colbert_small

Multi-vector (late-interaction) embeddings of **BEIR nfcorpus** (`beir/nfcorpus/test`), encoded with
**[lightonai/answerai-colbert-small-v1](https://huggingface.co/lightonai/answerai-colbert-small-v1)** at revision `e507cd12947a2b4b52201d150967df3c19a90590`.

**Source data:** [ir_datasets](https://ir-datasets.com/beir.html#beir/nfcorpus/test) `beir/nfcorpus/test` (ir_datasets 0.6.3), which downloads [nfcorpus.zip](https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/nfcorpus.zip) (md5 `a89dba18a62ef92f7d323ec890a0d38d`). BEIR also publishes this corpus on the Hub as [`BeIR/nfcorpus`](https://huggingface.co/datasets/BeIR/nfcorpus), whose card gives this dataset's license; the data here was loaded through ir_datasets, not from that repo. Document, query and qrel ids are the source's own ids,
unchanged.

Every document is one variable-length set of 96-d vectors; every query is one variable-length
set of 96-d vectors. Documents and queries are stored at different precisions (fp16 and fp32
respectively), see [Encoding](#encoding).

## Files

| file | dtype | shape | contents |
|---|---|---|---|
| `documents.npy` | float16 (`<f2`) | `[855,519, 96]` | every document vector, concatenated document by document (156.7 MiB) |
| `doclens.npy` | int32 | `[3,633]` | vectors per document; `cumsum` gives offsets |
| `token_ids.npy` | uint32 | `[855,519]` | tokenizer id of each `documents.npy` row, 1:1 |
| `doc_ids.npy` | `<U8` | `[3,633]` | original document ids |
| `queries.npy` | float32 (`<f4`) | `[323, 32, 96]` | query vectors, zero-padded at the end (3.8 MiB) |
| `query_lens.npy` | int32 | `[323]` | true vectors per query, before padding |
| `queries_ids.npy` | `<U10` | `[323]` | original query ids |
| `qrels.test.tsv` | text | 12,334 rows | TREC qrels, `qid \t 0 \t docid \t relevance`, no header |
| `gt_top1000.tsv` | text | 323,000 rows | exact MaxSim top-1000, see below |
| `gt_top100.tsv` | text | 32,300 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 | 3,633 |
| document vectors | 855,519 |
| vectors per document (min / median / mean / max) | 27 / 249 / 235.5 / 287 |
| queries | 323 |
| vectors per query (min / median / mean / max) | 32 / 32 / 32.0 / 32 |
| queries with at least one qrel | 323 |
| qrels rows | 12,334 |
| embedding dimension | 96 |

## Encoding

| | |
|---|---|
| model | [lightonai/answerai-colbert-small-v1](https://huggingface.co/lightonai/answerai-colbert-small-v1) |
| model revision | `e507cd12947a2b4b52201d150967df3c19a90590` |
| library | sentence-transformers 6.1.0 `MultiVectorEncoder` (transformers 5.17.0, torch 2.13.0+cu126) |
| 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 | 300 tokens (the checkpoint's `document_length`), applied before the skiplist. 2,473 of 3,633 documents (68%) were longer and were cut to it; longest here 287 vectors |
| query truncation | `query_length` unset; fixed query expansion: every query is padded to 32 tokens with the tokenizer's mask token, which are not attended to, and those expansion vectors are kept |
| document skiplist | 32 words removed: ['!', '"', '#', '$', '%', '&', "'", '(', ')', '*', '+', ',', '-', '.', '/', ':', ';', '<', '=', '>', '?', '@', '[', '\\', ']', '^', '_', '`', '{', '|', '}', '~'] |
| document input | `title + "\n\n" + text` when the corpus has a title, else `text`; stripped |
| 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 | tokenizer id of each kept document token (after the skiplist above), aligned 1:1 with `documents.npy` |

## 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 |
|---|---|---|---|---|---|
| 0.3683 | 0.5874 | 0.6780 | 0.3140 | 0.6283 | 0.1931 |

## Loading

```python
import numpy as np

documents = np.load("documents.npy", mmap_mode="r")      # [n_tokens, 96] 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], 96]

queries = np.load("queries.npy")                         # [n_queries, 32, 96] 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], 96]

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 (855519, 96)
- ✅ doclens.npy dtype/shape — <i4 (3633,)
- ✅ doc_ids.npy is a string array — <U8 (3633,)
- ✅ queries.npy dtype/shape — <f4 (323, 32, 96)
- ✅ query_lens.npy dtype/shape — <i4 (323,)
- ✅ queries_ids.npy is a string array — <U10 (323,)
- ✅ sum(doclens) == n_tokens — 855519 vs 855519
- ✅ no empty documents — min doclen 27
- ✅ len(doc_ids) == len(doclens) == corpus size — 3633, 3633, 3633
- ✅ doc_ids unique
- ✅ query arrays aligned — 323, 323, 323
- ✅ doc and query dim agree — 96 / 96
- ✅ token_ids.npy dtype/shape — <u4 (855519,)
- ✅ document vectors unit-norm (100k sample) — norm range [0.9994, 1.0005]
- ✅ query vectors unit-norm — norm range [1.000000, 1.000000]
- ✅ all vectors finite
- ✅ gt_top100.tsv has k rows per query — 32300 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 — 323000 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 | 2026-09-25 |
| hardware | Tesla V100S-PCIE-32GB |
| 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: 385 rows differ from the previous file, all of them documents with identical scores listed in a different order (3 tied pairs at rank 100 swapped in or out) |