Japanese-focused Gemma-3-12B tune
Hub tags carry gemma-3, Japanese, and ja markers with a google/gemma-3-12b-it source tag, and the model card frames the work as improving the base model's Japanese performance.
Open Source Model Profile · AXCXEPT
EZO2.5-gemma-3-12b-it-Preview
EZO2.5-gemma-3-12b-it-Preview is a 12.19B-parameter Gemma3 Japanese fine-tune from AXCXEPT. Its card documents EZO training of gemma-3-12b-it on 3,000 samples.
- Publisher
- AXCXEPT
- Task
- image-text-to-text
- Model type
- gemma3
- License
- gemma
- Library
- transformers
- Publication status
- Accepted · not indexed
Model overview
EZO2.5-gemma-3-12b-it-Preview is published by AXCXEPT as a Gemma3 image-text-to-text fine-tune. The captured configuration identifies Gemma3ForConditionalGeneration and Safetensors metadata reports 12,187,325,040 parameters, or about 12.19B. The model card describes a Japanese-focused tune of google/gemma-3-12b-it using the publisher's EZO method with GRPO and PPO concepts.
Recorded capabilities
EZO with GRPO and PPO concepts
According to the model card, the publisher mixed GRPO and PPO concepts into its EZO training method, described as still in research phase and needing automation and ablation work.
Documented vLLM serving pattern
The model card states the preview runs on a single A40 GPU and shows vLLM serving with max-model-len 32768 alongside an OpenAI-compatible client example.
Use cases in the source record
- Japanese-language text and chat experiments building on the documented gemma-3-12b-it Japanese improvement goal.
- Single-GPU vLLM serving experiments following the publisher's A40 and max-model-len 32768 example, subject to local validation.
Limitations and unknowns
- Japanese benchmark improvement is claimed without numeric scores in the captured evidence; no evaluation table was extracted.
- Provider state is historical snapshot data and should be refreshed before being presented as current.
- No context-window value was extracted beyond the serving example's 32768 max-model-len setting, which is a runtime argument rather than an architecture claim.
- The publisher describes the EZO method as research-stage and not yet automated or ablated.
Source and provenance
Source: AXCXEPT/EZO2.5-gemma-3-12b-it-Preview
Captured: Unknown. Processed: 2026-09-07T19:35:02.023277+00:00.
AXCXEPT/EZO2.5-gemma-3-12b-it-Preview Model Details 昨今登場したLLM自身の力を自力で向上させる「GRPO」や「PPO」の概念を、 弊社で開発した「EZO」というトレーニング手法にミックスすることで、 3,000件のデータセット、2時間×H200×8台のトレーニングで、Japanese MT Benchおよび、Elyza Tasks100におけるベースモデルの日本語性能を向上させることに成功したモデルです。 本トレーニング手法は、まだ研究段階にあり手法の自動化や、アブレーションが必要なステータスではあるものの、複雑かつ非常に時間がかかるGRPO/PPOといった強化学習方法を、 低予算でも実現できる大体の手段となりえると考えています。 By integrating the recently introduced concepts of “GRPO” and “PPO” — which enable LLMs to autonomously improve their own capabilities — into our proprietary training method “EZO,” we successfully enhanced the Japanese performance of the base model on both Japanese MT Bench and Elyza Tasks100…
F001F002F003F004F005F006F007F009F010F012F013F014F015F016