Sarashina2.2 provides 0.5‑B, 1‑B, and 3‑B language models trained by SB Intuitions through a three‑phase pipeline and synthetic data, achieving top Japanese QA, math, and coding scores while offering pre‑trained weights that are not instruction‑tuned and may produce biased outputs.
Description in other languages
Statistics
Statistics as of Nov 16, 2026
Hugging Face downloads (last 30 days)
1,000
Hugging Face likes
16
Listed
Dec 15, 2025
Downloads (30-day window) over time
Show as a table
| Month | DL (30d) |
|---|---|
| Dec 2025 | 37,000 |
| Jan 2026 | — |
| Feb 2026 | — |
| Mar 2026 | — |
| Apr 2026 | 1,000 |
| May 2026 | 1,000 |
| Jun 2026 | 3,000 |
| Jul 2026 | 2,000 |
| Aug 2026 | 1,000 |
| Sep 2026 | 2,000 |
| Oct 2026 | 4,000 |
| Nov 2026 | 1,000 |
Downloads over the 30 days before the last snapshot of each month.
Popular in the same category
- sarashina2.2-3bsbintuitions2kA 3‑billion‑parameter Japanese‑focused LLM, Sarashina2.2‑3B, trained in three phases—including synthetic math/coding data and fine‑tuning on small application data—outperforms larger models on Japanese QA, math and coding benchmarks, but lacks instruction tuning and can produce biased or nonsensical outputs.
- bert-base-japanese-char-v3tohoku-nlp130kJapanese‑language BERT‑Base (12 layers, 768‑dim, 12 heads) pretrained with Unidic‑based word‑level plus character‑level tokenization and whole‑word masking on CC‑100 and 2023 Wikipedia, producing a 7,027‑token vocabulary.
- bert-large-japanese-v2tohoku-nlp66kJapanese‑BERT‑Large trained on CC‑100 and Wikipedia, using Unidic‑lite word‑level tokenization with WordPiece subwords and whole‑word masking (24 layers, 1024‑dim hidden, 16 heads, 32k vocab), with pretraining code on cl‑tohoku/bert‑japanese.
- bert-base-japanese-v3tohoku-nlp66kJapanese BERT‑base (12 layers, 768‑dim hidden, 12 heads, 32 k vocab) pretrained with whole‑word masking on CC‑100 and 2023‑Jan Wikipedia, using Unidic 2.1.2 word‑level tokenization plus WordPiece, in 2 M training steps.
- kana-whispersbintuitions42kA fine‑tuned Whisper large‑v3‑turbo model that transcribes Japanese speech into katakana, serving as the ASR component of the Joyo Kanji Yomi Benchmark within the Sarashina2.2‑TTS project and powering the Kana CER Usage With Transformers pipeline.
- t5-base-japanese-v1.1sonoisa17kA Japanese T5‑v1.1 model pretrained on ≈100 GB of Wikipedia and OSCAR CC‑100 data (mixed 10:1 for SentencePiece with byte‑fallback), requiring fine‑tuning for downstream tasks, includes transfer‑learning sample code, notes potential bias in outputs, and is licensed CC‑BY‑SA 4.0.
Resources in similar research fields
japanese-nli-modelCyberAgentAILab6This repository provides the code for Japanese NLI model, a fine-tuned masked language model.GitHubShared: QA & reasoning
jrte-corpus_examplemegagonlabs3Example codes for the Japanese Realistic Textual Entailment Corpus.GitHubShared: QA & reasoning- JCommonsenseQAsbintuitions1kJCommonsenseQA is a Japanese, multiple‑choice dataset for commonsense reasoning—an adaptation of CommonsenseQA—licensed under CC BY‑SA 4.0 and cited as doi:10.5715/jnlp.30.63.HF datasetShared: QA & reasoning
- embhpprc1kA catalog of Japanese and multilingual QA, NLI and paraphrase datasets, detailing each dataset’s retrieval or QA tasks and its license (Apache 2.0, CC‑BY‑SA/CC‑BY, MIT, etc.).HF datasetShared: QA & reasoning
- JaQuADSkelterLabsInc916JaQuAD is a 2022 Japanese QA dataset of 39,696 SQuAD‑style extractive pairs from Wikipedia, totaling 73.2 MB, that achieves 78.92 % F1 (63.38 % EM) when fine‑tuned with BERT‑Japanese.HF datasetShared: QA & reasoning