Shisa V2
Shisa V2 is a family of bilingual Japanese and English (JA/EN) general-purpose chat models trained by
Shisa.AI. These models aim to excel in Japanese language tasks while retaining robust English capabilities.
Since our initial
Shisa 7B releases, the baseline Japanese capabilities of open-weight language models have significantly improved. New models have more Japanese pre-training tokens, higher
JA tokenizer efficiency, and better quality Japanese outputs overall. As such, for Shisa V2 we've eschewed both tokenizer extension and costly continued pre-training and have focused entirely on optimizing post-training. We've significantly expanded and refined the synthetic-data driven approach that was pioneered with our original
Shisa 7B models, and have achieved substantial performance gains.
Model Family Overview
The Shisa V2 family comprises a range of models from 7B to 70B parameters in size:
These Shisa V2 models were all trained using the same datasets and training recipes, except for scaling the learning rate based on model size and modifying the global batch size for the 70B model.
While most of our development and tuning was done on the Llama 3.1 8B model, we did some cross-validation during this process and we're pleased that our final recipe has shown robust scaling, improving Japanese language performance across all model sizes evaluated. We've prioritized releasing the highest-quality openly-licensed (Apache 2.0 and MIT) models in each class size.
Performance
All Shisa V2 models demonstrate improved Japanese output quality compared to their respective base models:
| Model Name | JA AVG | EN AVG | Shaberi AVG | ELYZA 100 | JA MT Bench | Rakuda | Tengu | llm-jp-eval | shisa-jp-ifeval | shisa-jp-rp-bench | shisa-jp-tl-bench | MixEval | LiveBench | IFEval | EvalPlus |
|---|
| shisa-ai/shisa-v2-qwen2.5-32b | 76.97 | 67.41 | 8.63 | 8.59 | 8.88 | 9.04 | 8.00 | 0.66 | 0.40 | 4.71 | 7.07 | 0.53 | 49.9 | 0.82 | 0.84 |
| Qwen/Qwen2.5-32B-Instruct | 66.79 | 66.67 | 8.41 | 8.59 | 8.64 | 8.61 | 7.80 | 0.66 | 0.43 | 4.58 | 5.01 | 0.53 | 49.7 | 0.82 | 0.82 |
The Shisa V2 models perform well against other models in their respective class sizes.
| License | Model Name | JA AVG | EN AVG | Shaberi AVG | ELYZA 100 | JA MT Bench | Rakuda | Tengu | llm-jp-eval | shisa-jp-ifeval | shisa-jp-rp-bench | shisa-jp-tl-bench | MixEval | LiveBench | IFEval | EvalPlus |
|---|
| Gemma | google/gemma-3-27b-it | 80.05 | 65.95 | 8.53 | 8.53 | 8.71 | 8.85 | 8.03 | 0.65 | 0.57 | 4.67 | 8.38 | 0.48 | 51.8 | 0.85 | 0.79 |
| Apache 2.0 | shisa-ai/shisa-v2-qwen2.5-32b | 76.97 | 67.41 | 8.63 | 8.59 | 8.88 | 9.04 | 8.00 | 0.66 | 0.40 | 4.71 | 7.07 | 0.53 | 49.9 | 0.82 | 0.84 |
| Apache 2.0 | abeja/ABEJA-Qwen2.5-32b-Japanese-v0.1 | 74.14 | 65.70 | 8.66 | 8.65 | 8.78 | 9.07 | 8.14 | 0.66 | 0.37 | 4.12 | 6.20 | 0.55 | 46.20 | 0.83 | 0.79 |
| Apache 2.0 | Qwen/Qwen2.5-32B-Instruct | 73.35 | 66.67 | 8.41 | 8.59 | 8.64 | 8.61 | 7.80 | 0.66 | 0.43 | 4.58 | 5.01 | 0.53 | 49.7 | 0.82 | 0.82 |
| Apache 2.0 | mistralai/Mistral-Small-3.1-24B-Instruct-2503 | 72.03 | 65.15 | 8.51 | 8.56 | 8.63 | 9.12 | 7.74 | 0.63 | 0.36 | 4.59 | 4.49 | 0.55 | 44.90 | 0.81 | 0.80 |
Testing Notes
Japanese functional tests were conducted using the
shisa-ai/shaberi fork of the
LightBlue Shaberi evaluation harness. Shaberi ratings were performed with a
PoLL (LLM Jury) consisting of:
The results were statistically validated to be comparable to both gpt-4-1106-preview and human-reviewed "gold standard" ratings.
Dynamic RoPE extension was utilized when necessary for testing models with context windows smaller than 8K tokens. All tests were performed using recent versions of
vLLM or
SGLang.
We developed a custom "multieval" harness to automate our model evaluations. Standard benchmarks include:
New Japanese Benchmarks
Over the course of model development, we also created several new evaluations to help us measure performance on important Japanese downstream tasks:
- shisa-jp-ifeval: Inspired by IFEval, but evaluating instruction-following abilities specific to Japanese grammar and linguistics (closed form)
- shisa-jp-rp-bench: Assessing performance on Japanese role-play and character/persona-based multi-turn conversations based on Aratako's Japanese-RP-Bench (LLM judge)
- shisa-jp-tl-bench: Testing Japanese-English translation proficiency (LLM judge, BTL pairwise comparison with logistic transformation scoring)
We believe these benchmarks will be generally useful and plan to open-source them in the near future to support the Japanese LLM research community.
Usage
All Shisa V2 models inherit the
chat templates of their respective base models and have been tested and validated for proper inference with both
vLLM and
SGLang.
Running sampler sweeps, we found the models operate well across a variety of temperatures in most settings. For translation tasks specifically, we recommend a lower temperatures (0.2) to increase accuracy. For role-play and creative tasks, a higher temp (eg 1.0) seems to give good results. To prevent cross-lingual token leakage we recommend a top_p of 0.9 or min_p of 0.1.
No additional safety alignment has been done on these models, so they will largely inherit the base models' biases and safety profiles.
Datasets
Our supervised fine-tuning (SFT) stage dataset consists of approximately 360K samples totaling roughly 420M Llama 3 tokens:
- shisa-ai/shisa-v2-sharegpt
- This is a filtered, regenerated and resampled version of the original Shisa V1 augmxnt/ultra-orca-boros-en-ja-v1 dataset
- This was the backbone of our Shisa V2 training and it proved to be an extremely robust dataset, out-performing all existing mixes/additions (Tulu, Olmo, Rewild, various Magpie sets, etc.) - if you need a JA/EN dataset, we believe this new version is among the best currently available
- shisa-ai/rewild-set-deepseek-subset
- A filtered version of Rewild (WildChat) prompts translated into Japanese, with responses generated by DeepSeek-V3-0324
- shisa-ai/magpie-ultra-set
- Japanese generations based on argilla/magpie-ultra-v1.0
- shisa-ai/magpie-advanced-questions-set
- Magpie-generated questions about advanced college-level topics across a variety of academic fields
- shisa-ai/japan-magpie-set
- Magpie-generated questions about Japan's economy and history as well as cultural and business practices
- shisa-ai/shisa-v2-roleplaying-sft
- Synthetically-generated roleplaying data featuring a wide variety of characters, situations, and genres
- shisa-ai/translation_expanded_master_set_filtered
- A synthetic dataset involving a wide range of translation tasks, including essays, conversations, and fiction
- shisa-ai/shisa-v2-instruction-following-sft
- An instruction following dataset based on prompts from (Aratako/Magpie-Tanuki-8B-annotated-96k) and a list of instruction-following constraints
Our final DPO mix is 113K samples totaling approximately 115M Llama 3 tokens:
- shisa-ai/deepseekv3-ultrafeedback-armorm-dpo
- This is a version of princeton-nlp/gemma2-ultrafeedback-armorm with
chosen responses regenerated by DeepSeek-V3-0324
- Surprisingly, we found that using this relatively small DPO alignment set in English-only outperformed both JA/EN DPO sets and also much larger sets like the Tulu 3 preference mixture
- shisa-ai/shisa-v2-roleplaying-dpo
- A DPO variant of the roleplaying-sft set that uses an UltraFeedback-style rating system
- shisa-ai/translation-no-extra-text-dpo-dataset
- A DPO set that aims to reduce the tendency of models to output extraneous explanatory text for translations when not wanted
- shisa-ai/shisa-v2-instruction-following-dpo
- A DPO variant of the instruction-following-sft set to further enhance instruction-following performance
- shisa-ai/politeness-dpo-set
- A set to allow for greater controllability of speaking style for Japanese responses
Training
We trained over 200 models to empirically test a wide range of variables. Beyond hyper-parameter and data-mix testing, we also ran numerous tests on data ordering, multilingual-specific ordering, curriculum learning, multi-stage training, various forms of self-play, preference tuning, and some of the latest RL/verifiable reward techniques.
A full discussion of these learnings is out of scope here, but we will be updating the
shisa-v2 wiki and the
Shisa.AI website with forthcoming writeups.
Most of our training was done on a small AWS Sagemaker-deployed 4-node H100 slurm cluster. Training was mostly done with
Axolotl with
DeepSpeed and
Liger Kernels. The Phi 4 and Llama 3.3 70B versions of Shisa V2 were trained with
OpenRLHF. Our training logs are
publicly available on Weights and Biases.
Credits
The Shisa V2 models were developed by
Leonard Lin and Adam Lensenmayer (
Shisa.AI).
Compute was provided by
Ubitus K.K. and
METI GENIAC.
Thanks to
Meta Llama,
Microsoft Research,
Mistral AI, and
Qwen Team for providing their models to the open source community,
Unsloth for their
llamafied conversion of Phi-4, the Tulu team, whose detailed writeups and fast responses to our questions were very helpful, and
Chanvichet Vong of the Axolotl team for his tireless work in the Axolotl Discord.
We also extend our thanks to all open source AI developers and researchers - without their publicly shared research, tooling, and datasets, none of our work would be possible. We hope that our own contributions will further support the broader community.
A special thanks to
Jon Durbin for his work on Shisa V1.
For more details on our development and insights, please visit the
Shisa V2 Github repository and the
Shisa.AI website.
1: Per the Llama Community License Agreements, the official names of the Llama-based models are "Llama 3.1 shisa-v2-llama3.1-8b" and "Llama 3.3 shisa-v2-llama3.3-70b"
