LLM-jp-3 172B Instruct3
LLM-jp-3 172B Instruct3 adds supervised fine-tuning and DPO to the final 172B model. Its custom license differs from the gated base and Instruct2 releases.
Research release
Preference alignment follows instruction tuning
The card describes synthetic preference data used to improve helpfulness and harmlessness. SFT includes Japanese human-written instructions, safety examples, synthetic conversation, English instruction data and Japanese-English coding data.LLM-jp-3 172B Instruct3 official model card
Review attribution in downstream releases
Practical guidance. The broader redistribution permission makes this release more suitable than the base model for distributing an adapted checkpoint. Keep the required attribution with the released product and model files.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | llm-jp-3-172b-instruct3 |
| Context window | 4,096 tokens Maximum sequence length stated in the model card. LLM-jp-3 172B Instruct3 official model card |
| Parameters | 172,392,173,568 total.LLM-jp-3 172B Instruct3 official model card |
| Architecture | 96 layers; hidden size 12,288; 96 attention heads.LLM-jp-3 172B Instruct3 official model card |
| License | Custom agreement permits commercial use, modification and redistribution. It requires the agreement, copyright notice and a Built with llm-jp-3-172b-instruct3 attribution, with additional use restrictions.LLM-jp-3 172B Instruct3 license agreement |
Related models
- LLM-jp-3 172B Base
LLM-jp / NII LLMC · LLM-jp-3 172B
The final 2.1-trillion-token 172B base checkpoint uses application-based access and custom terms.
Launch date not confirmedRestricted accessDownloadable - LLM-jp-3 172B Instruct2
LLM-jp / NII LLMC · LLM-jp-3 172B
A gated 172B supervised instruction release for comparison with preference-aligned Instruct3.
Launch date not confirmedRestricted accessDownloadable