NanoMIRACL

A tiny, evaluation-ready slice of MIRACL that mirrors the spirit of NanoBEIR: same task, same style, but dramatically smaller so you can iterate and benchmark in minutes instead of hours.

Evaluation can be performed during and after training by integrating with Sentence Transformer's Evaluation module (InformationRetrievalEvaluator).

NanoMIRACL Evaluation (NDCG@10)

Model	Avg	ar	bn	de	en	es	fa	fi	fr	hi	id	ja	ko	ru	sw	te	th	yo	zh
multilingual-e5-small	0.7157	0.7421	0.7233	0.6519	0.6976	0.7241	0.7224	0.7761	0.6394	0.6801	0.6171	0.6941	0.6713	0.7195	0.7206	0.9779	0.8836	0.6125	0.6287
multilingual-e5-large	0.7766	0.8108	0.8049	0.7951	0.7260	0.7996	0.7426	0.8346	0.6662	0.7751	0.6147	0.7398	0.7058	0.8020	0.7500	0.9779	0.9033	0.8109	0.7192
e5-small-v2	0.2638	0.0067	0.0000	0.4828	0.7220	0.5946	0.0126	0.5333	0.5068	0.0271	0.3137	0.1526	0.0286	0.1588	0.4252	0.0200	0.0200	0.6306	0.1137
e5-large-v2	0.2970	0.0460	0.0260	0.5281	0.7494	0.6416	0.0326	0.5552	0.5380	0.0571	0.3494	0.2756	0.0858	0.3536	0.3917	0.0200	0.0086	0.5641	0.1232
bge-m3	0.7880	0.8257	0.8440	0.7307	0.7386	0.7516	0.7794	0.7863	0.7252	0.7408	0.6619	0.8044	0.7337	0.7968	0.7550	0.9926	0.9084	0.8547	0.7549
gte-multilingual-base	0.7430	0.7817	0.7264	0.6918	0.7101	0.7107	0.7328	0.8067	0.7048	0.7067	0.6281	0.7533	0.6480	0.7525	0.6927	0.9900	0.8308	0.7807	0.7269
nomic-embed-text-v2-moe	0.7492	0.7903	0.8084	0.6693	0.7097	0.7250	0.7431	0.7979	0.6826	0.7349	0.6360	0.7186	0.6688	0.7759	0.6963	0.9926	0.8860	0.7637	0.6874
paraphrase-multilingual-MiniLM-L12-v2	0.4148	0.5184	0.0334	0.5161	0.6042	0.6499	0.4048	0.5234	0.5605	0.3865	0.4736	0.3938	0.4131	0.4892	0.1722	0.0200	0.5489	0.2885	0.4693

Notes:

• The above results were computed with ./nano_eval.py.
• E5 models were evaluated with --query-prompt "query: " and --corpus-prompt "passage: ".
• nomic-ai/nomic-embed-text-v2-moe was evaluated with --query-prompt "search_query: " and --corpus-prompt "search_document: ".
• Some models (e.g., BAAI/bge-m3, Alibaba-NLP/gte-multilingual-base, nomic-ai/nomic-embed-text-v2-moe) require --trust-remote-code.
• This benchmark section is a draft and may be refined as scripts/metadata are finalized.

What this dataset is

• A collection of 18 language subsets (corpus, queries, qrels) published on the Hugging Face Hub under hotchpotch/NanoMIRACL.
• Each subset contains 50 dev queries and a corpus of 10,000 documents.
• Queries are sampled from hotchpotch/miracl-hf-unified (derived from MIRACL), keeping one positive per query and all hard negatives, then filling the corpus to 10k with random documents.
• Sampling is deterministic (seed=42).
• License: Other (see MIRACL and upstream repository licenses).

Subset names

• Split names:
  • NanoMIRACL-ar
  • NanoMIRACL-bn
  • NanoMIRACL-de
  • NanoMIRACL-en
  • NanoMIRACL-es
  • NanoMIRACL-fa
  • NanoMIRACL-fi
  • NanoMIRACL-fr
  • NanoMIRACL-hi
  • NanoMIRACL-id
  • NanoMIRACL-ja
  • NanoMIRACL-ko
  • NanoMIRACL-ru
  • NanoMIRACL-sw
  • NanoMIRACL-te
  • NanoMIRACL-th
  • NanoMIRACL-yo
  • NanoMIRACL-zh
• Config names: corpus, queries, qrels

Usage

from datasets import load_dataset

split = "NanoMIRACL-ja"
queries = load_dataset("hotchpotch/NanoMIRACL", "queries", split=split)
corpus  = load_dataset("hotchpotch/NanoMIRACL", "corpus",  split=split)
qrels   = load_dataset("hotchpotch/NanoMIRACL", "qrels",   split=split)

print(queries[0]["text"])

Example eval code

python ./nano_eval.py   --model-path intfloat/multilingual-e5-small   --query-prompt "query: "   --corpus-prompt "passage: "

For models that require trust_remote_code, add --trust-remote-code (e.g., BAAI/bge-m3).

Why Nano?

• Fast eval loops: 50 queries × 10k docs fits comfortably on a single GPU/CPU run.
• Reproducible: deterministic sampling and stable IDs.
• Drop-in: BEIR/NanoBEIR-style schemas, so existing IR loaders need minimal tweaks.

Upstream sources

• Original data: MIRACL — Multilingual Information Retrieval Across a Continuum of Languages.
• Base dataset: hotchpotch/miracl-hf-unified (Hugging Face Hub).
• Inspiration: NanoBEIR (lightweight evaluation subsets).

License

Other. This dataset is derived from MIRACL and ultimately from open-source sources. Please respect original repository licenses and attribution requirements.

Author

• Yuichi Tateno