README.md

---
configs:
- config_name: bm25
  data_files:
  - split: NanoArguAna
    path: bm25/NanoArguAna.parquet
  - split: NanoClimateFEVER
    path: bm25/NanoClimateFEVER.parquet
  - split: NanoDBPedia
    path: bm25/NanoDBPedia.parquet
  - split: NanoFEVER
    path: bm25/NanoFEVER.parquet
  - split: NanoFiQA2018
    path: bm25/NanoFiQA2018.parquet
  - split: NanoHotpotQA
    path: bm25/NanoHotpotQA.parquet
  - split: NanoMSMARCO
    path: bm25/NanoMSMARCO.parquet
  - split: NanoNFCorpus
    path: bm25/NanoNFCorpus.parquet
  - split: NanoNQ
    path: bm25/NanoNQ.parquet
  - split: NanoQuoraRetrieval
    path: bm25/NanoQuoraRetrieval.parquet
  - split: NanoSCIDOCS
    path: bm25/NanoSCIDOCS.parquet
  - split: NanoSciFact
    path: bm25/NanoSciFact.parquet
  - split: NanoTouche2020
    path: bm25/NanoTouche2020.parquet
- config_name: corpus
  data_files:
  - split: NanoArguAna
    path: corpus/NanoArguAna.parquet
  - split: NanoClimateFEVER
    path: corpus/NanoClimateFEVER.parquet
  - split: NanoDBPedia
    path: corpus/NanoDBPedia.parquet
  - split: NanoFEVER
    path: corpus/NanoFEVER.parquet
  - split: NanoFiQA2018
    path: corpus/NanoFiQA2018.parquet
  - split: NanoHotpotQA
    path: corpus/NanoHotpotQA.parquet
  - split: NanoMSMARCO
    path: corpus/NanoMSMARCO.parquet
  - split: NanoNFCorpus
    path: corpus/NanoNFCorpus.parquet
  - split: NanoNQ
    path: corpus/NanoNQ.parquet
  - split: NanoQuoraRetrieval
    path: corpus/NanoQuoraRetrieval.parquet
  - split: NanoSCIDOCS
    path: corpus/NanoSCIDOCS.parquet
  - split: NanoSciFact
    path: corpus/NanoSciFact.parquet
  - split: NanoTouche2020
    path: corpus/NanoTouche2020.parquet
- config_name: qrels
  data_files:
  - split: NanoArguAna
    path: qrels/NanoArguAna.parquet
  - split: NanoClimateFEVER
    path: qrels/NanoClimateFEVER.parquet
  - split: NanoDBPedia
    path: qrels/NanoDBPedia.parquet
  - split: NanoFEVER
    path: qrels/NanoFEVER.parquet
  - split: NanoFiQA2018
    path: qrels/NanoFiQA2018.parquet
  - split: NanoHotpotQA
    path: qrels/NanoHotpotQA.parquet
  - split: NanoMSMARCO
    path: qrels/NanoMSMARCO.parquet
  - split: NanoNFCorpus
    path: qrels/NanoNFCorpus.parquet
  - split: NanoNQ
    path: qrels/NanoNQ.parquet
  - split: NanoQuoraRetrieval
    path: qrels/NanoQuoraRetrieval.parquet
  - split: NanoSCIDOCS
    path: qrels/NanoSCIDOCS.parquet
  - split: NanoSciFact
    path: qrels/NanoSciFact.parquet
  - split: NanoTouche2020
    path: qrels/NanoTouche2020.parquet
- config_name: queries
  data_files:
  - split: NanoArguAna
    path: queries/NanoArguAna.parquet
  - split: NanoClimateFEVER
    path: queries/NanoClimateFEVER.parquet
  - split: NanoDBPedia
    path: queries/NanoDBPedia.parquet
  - split: NanoFEVER
    path: queries/NanoFEVER.parquet
  - split: NanoFiQA2018
    path: queries/NanoFiQA2018.parquet
  - split: NanoHotpotQA
    path: queries/NanoHotpotQA.parquet
  - split: NanoMSMARCO
    path: queries/NanoMSMARCO.parquet
  - split: NanoNFCorpus
    path: queries/NanoNFCorpus.parquet
  - split: NanoNQ
    path: queries/NanoNQ.parquet
  - split: NanoQuoraRetrieval
    path: queries/NanoQuoraRetrieval.parquet
  - split: NanoSCIDOCS
    path: queries/NanoSCIDOCS.parquet
  - split: NanoSciFact
    path: queries/NanoSciFact.parquet
  - split: NanoTouche2020
    path: queries/NanoTouche2020.parquet
  default: true
language:
- multilingual
tags:
- information-retrieval
- retrieval
- nano
- bm25
- hakari-bench
---

# NanoBEIR-th

This dataset is a Nano-style retrieval dataset. Nano-series evaluation can
be run easily with [HAKARI-Bench](https://github.com/hotchpotch/hakari-bench).

NanoBEIR-th is derived from MNanoBEIR / NanoBEIR. It follows the
Hugging Face Datasets layout convention used by
[sentence-transformers/NanoBEIR-en](https://huggingface.co/datasets/sentence-transformers/NanoBEIR-en):
each Nano split has separate `corpus`, `queries`, and `qrels` tables, and BM25
candidates are provided separately in a `bm25` table. This layout follows
the NanoBEIR-style evaluation approach summarized in
[NanoBEIR](https://huggingface.co/blog/sionic-ai/eval-sionic-nano-beir).

NanoBEIR-th is the Thai language-specific component of MNanoBEIR. It groups compact BEIR-derived retrieval tasks for efficient evaluation of document ranking in that language.



## Source Links

- Final dataset: [hakari-bench/NanoBEIR-th](https://huggingface.co/datasets/hakari-bench/NanoBEIR-th)

## Data Layout

This dataset uses four Hugging Face Datasets configs:

- `corpus`: documents with `_id` and `text`
- `queries`: queries with `_id` and `text`
- `qrels`: positive relevance labels with `query-id` and `corpus-id`
- `bm25`: BM25 candidate lists with `query-id` and `corpus-ids`

Each config has the same Nano split names.

The `qrels` config is positive-only. Source rows with `score <= 0` are treated
as non-relevant or hard-negative annotations and are not included in `qrels`.
When the source provides such rows, their documents are preferentially used as
hard negatives in the `corpus` config before generic corpus-fill documents.
Source hard negatives are sampled deterministically with query round-robin so
one query's negative pool does not dominate the corpus.

## Split Statistics

Length statistics are computed with `len(str(text))` over the `queries` and
`corpus` tables. `std` is the population standard deviation over the rows in
each split.

| Nano split | Queries | Corpus | Qrels | Query avg | Query std | Query median | Query p25 | Query p75 | Doc avg | Doc std | Doc median | Doc p25 | Doc p75 |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| NanoArguAna | 50 | 3635 | 50 | 820.6 | 133.1 | 881.5 | 772.2 | 913.0 | 860.1 | 467.1 | 771.0 | 535.0 | 1073.0 |
| NanoClimateFEVER | 50 | 3408 | 148 | 118.6 | 46.2 | 113.0 | 86.2 | 151.2 | 1395.4 | 732.3 | 1263.5 | 851.8 | 1799.2 |
| NanoDBPedia | 50 | 6045 | 1158 | 30.9 | 13.4 | 28.5 | 19.2 | 42.5 | 316.4 | 132.5 | 345.0 | 224.0 | 414.0 |
| NanoFEVER | 50 | 4996 | 57 | 46.9 | 15.8 | 45.5 | 36.8 | 57.0 | 1084.7 | 722.4 | 930.5 | 545.8 | 1455.2 |
| NanoFiQA2018 | 50 | 4598 | 123 | 55.2 | 19.5 | 51.5 | 39.5 | 70.0 | 779.2 | 667.4 | 574.5 | 314.0 | 999.8 |
| NanoHotpotQA | 50 | 5090 | 100 | 79.7 | 27.7 | 74.0 | 57.5 | 99.8 | 330.7 | 239.4 | 287.0 | 151.0 | 449.0 |
| NanoMSMARCO | 50 | 5043 | 50 | 32.1 | 13.6 | 29.0 | 25.2 | 38.5 | 293.9 | 115.0 | 266.0 | 221.0 | 339.0 |
| NanoNFCorpus | 50 | 2953 | 1651 | 22.6 | 13.3 | 20.5 | 10.2 | 31.0 | 1387.4 | 461.6 | 1409.0 | 1102.0 | 1647.0 |
| NanoNQ | 50 | 5035 | 57 | 40.8 | 12.5 | 39.0 | 33.0 | 44.8 | 473.6 | 424.0 | 402.0 | 161.0 | 673.0 |
| NanoQuoraRetrieval | 50 | 5046 | 70 | 46.9 | 16.9 | 42.0 | 37.2 | 56.0 | 53.7 | 27.5 | 47.0 | 36.0 | 62.0 |
| NanoSCIDOCS | 50 | 2210 | 244 | 69.1 | 19.4 | 68.5 | 57.0 | 82.0 | 820.4 | 528.2 | 814.5 | 510.2 | 1100.0 |
| NanoSciFact | 50 | 2919 | 56 | 92.7 | 39.1 | 88.5 | 57.5 | 117.8 | 1328.8 | 519.6 | 1245.0 | 978.5 | 1617.0 |
| NanoTouche2020 | 49 | 5745 | 932 | 46.3 | 16.0 | 42.0 | 37.0 | 55.0 | 1438.1 | 1332.4 | 886.0 | 287.0 | 2677.0 |

## Construction Steps

This dataset is constructed as follows.

1. Use MNanoBEIR / NanoBEIR as the upstream benchmark or dataset family.
2. Load source datasets from the `hakari-bench/NanoBEIR-th` corpus, queries, and qrels tables.
3. Source evaluation split policy: the NanoBEIR split set.
4. Create one Nano split for each selected source retrieval task.
5. Keep up to 200 eligible queries per Nano split.
6. Treat source relevance rows with `score > 0` as qrels-positive documents.
   If the source has no score column, treat its qrels as positive-only only when
   that is the source task convention.
7. Exclude source rows with `score <= 0` from `qrels`. When such rows are
   available for selected queries, use their documents as hard-negative corpus
   candidates before generic fill documents.
8. Include all qrels-positive documents for the selected queries.
9. Use the included corpus tables for each Nano split; no additional document resampling is performed.
10. Remove exact duplicate query text and document text within each split. If a
   removed document duplicate was referenced by qrels,
   the qrels row was removed.
11. Store corpus text as `title` plus body text when available.
12. Generate BM25 top-100 candidates with
    `wordseg:th` tokenization.
13. If a qrels-positive document is missing from the raw BM25 result, insert it
    into the final `bm25` candidate list by replacing a tail non-positive
    candidate.

The `qrels` config is positive-only. When needed for top-k reranking coverage, positive qrels are capped per query to the BM25 top-k.

The `bm25` candidate subset is generated from the included corpus for each split.

For top-100 reranking diagnostics, positive qrels are capped to at most 100 documents per query before BM25 positive forcing. This makes full relevant coverage possible for splits whose upstream qrels contain more than 100 positives for a single query.

## BM25 Subset Policy

The `bm25` config is a candidate subset for first-stage retrieval and reranking.
It is not a separate source dataset. Each row contains one query id and a ranked
list of up to 100 corpus ids.

BM25 candidates are generated from the selected corpus for each split. When a
qrels-positive document is not present in the raw BM25 top-100
results, the missing positive is forced into the final candidate list by
replacing a tail candidate that is not positive for that query. Candidate ids
are kept unique after replacement.

Concretely, each `bm25` row is produced by tokenizing the selected split corpus
and query texts with `wordseg:th`, ranking the corpus with BM25,
then writing the ranked corpus ids as `corpus-ids` for that query. The list is a
candidate subset for downstream evaluation, not a full-corpus ranking.

Source hard negatives, including documents referenced by source rows with
`score <= 0`, may appear in the selected corpus and can naturally appear in BM25
candidates. They are still non-relevant and are not listed in `qrels`.

When source hard negatives are available, the default corpus sampling policy is
query round-robin: group hard negatives by selected query, preserve source
rank/order within each query, add at most one new hard negative from each query
per pass, remove duplicate IDs and exact duplicate text, then fill any remaining
slots from source corpus order.

## Split Mapping

Each Nano split maps to one source retrieval task unless noted otherwise.

| Nano split | Source task | Source dataset | Queries | Corpus | Qrels |
|---|---|---|---:|---:|---:|
| NanoArguAna | NanoArguAna | [hakari-bench/NanoBEIR-th](https://huggingface.co/datasets/hakari-bench/NanoBEIR-th) | 50 | 3635 | 50 |
| NanoClimateFEVER | NanoClimateFEVER | [hakari-bench/NanoBEIR-th](https://huggingface.co/datasets/hakari-bench/NanoBEIR-th) | 50 | 3408 | 148 |
| NanoDBPedia | NanoDBPedia | [hakari-bench/NanoBEIR-th](https://huggingface.co/datasets/hakari-bench/NanoBEIR-th) | 50 | 6045 | 1158 |
| NanoFEVER | NanoFEVER | [hakari-bench/NanoBEIR-th](https://huggingface.co/datasets/hakari-bench/NanoBEIR-th) | 50 | 4996 | 57 |
| NanoFiQA2018 | NanoFiQA2018 | [hakari-bench/NanoBEIR-th](https://huggingface.co/datasets/hakari-bench/NanoBEIR-th) | 50 | 4598 | 123 |
| NanoHotpotQA | NanoHotpotQA | [hakari-bench/NanoBEIR-th](https://huggingface.co/datasets/hakari-bench/NanoBEIR-th) | 50 | 5090 | 100 |
| NanoMSMARCO | NanoMSMARCO | [hakari-bench/NanoBEIR-th](https://huggingface.co/datasets/hakari-bench/NanoBEIR-th) | 50 | 5043 | 50 |
| NanoNFCorpus | NanoNFCorpus | [hakari-bench/NanoBEIR-th](https://huggingface.co/datasets/hakari-bench/NanoBEIR-th) | 50 | 2953 | 1651 |
| NanoNQ | NanoNQ | [hakari-bench/NanoBEIR-th](https://huggingface.co/datasets/hakari-bench/NanoBEIR-th) | 50 | 5035 | 57 |
| NanoQuoraRetrieval | NanoQuoraRetrieval | [hakari-bench/NanoBEIR-th](https://huggingface.co/datasets/hakari-bench/NanoBEIR-th) | 50 | 5046 | 70 |
| NanoSCIDOCS | NanoSCIDOCS | [hakari-bench/NanoBEIR-th](https://huggingface.co/datasets/hakari-bench/NanoBEIR-th) | 50 | 2210 | 244 |
| NanoSciFact | NanoSciFact | [hakari-bench/NanoBEIR-th](https://huggingface.co/datasets/hakari-bench/NanoBEIR-th) | 50 | 2919 | 56 |
| NanoTouche2020 | NanoTouche2020 | [hakari-bench/NanoBEIR-th](https://huggingface.co/datasets/hakari-bench/NanoBEIR-th) | 49 | 5745 | 932 |

## BM25 nDCG@10

`nDCG@10` is computed from the included BM25 ranking against the included qrels.

Coverage is measured against included qrels at the same BM25 top-k used for reranking diagnostics. The included BM25 candidate subset has 100.00% query coverage and 100.00% relevant coverage for every split.

| Nano split | Tokenizer | Forced BM25 positives | Query cov | Relevant cov | BM25 nDCG@10 |
|---|---|---:|---:|---:|---:|
| NanoArguAna | wordseg:th | 3 | 100.00% | 100.00% | 0.4051 |
| NanoClimateFEVER | wordseg:th | 70 | 100.00% | 100.00% | 0.2368 |
| NanoDBPedia | wordseg:th | 403 | 100.00% | 100.00% | 0.5043 |
| NanoFEVER | wordseg:th | 3 | 100.00% | 100.00% | 0.7001 |
| NanoFiQA2018 | wordseg:th | 48 | 100.00% | 100.00% | 0.2726 |
| NanoHotpotQA | wordseg:th | 14 | 100.00% | 100.00% | 0.5523 |
| NanoMSMARCO | wordseg:th | 10 | 100.00% | 100.00% | 0.2907 |
| NanoNFCorpus | wordseg:th | 1397 | 100.00% | 100.00% | 0.3243 |
| NanoNQ | wordseg:th | 9 | 100.00% | 100.00% | 0.3191 |
| NanoQuoraRetrieval | wordseg:th | 2 | 100.00% | 100.00% | 0.7283 |
| NanoSCIDOCS | wordseg:th | 107 | 100.00% | 100.00% | 0.2641 |
| NanoSciFact | wordseg:th | 8 | 100.00% | 100.00% | 0.6334 |
| NanoTouche2020 | wordseg:th | 234 | 100.00% | 100.00% | 0.5108 |

## Skipped Tasks

No source tasks were skipped.

## License

NanoBEIR-th is a derived dataset. Users must comply with the licenses,
terms, and attribution requirements of the upstream source datasets.