Llama 3 Youko 8B Instruct
Youko 8B Instruct combines supervised fine-tuning with a chat vector from Meta's Llama 3 instruction model, followed by DPO.
Three stages add instruction behavior
The published merge adds half of the difference between Llama 3 8B Instruct and its base to rinna's SFT model. It excludes the embedding layer. The card also identifies direct preference optimization.Llama 3 Youko 8B Instruct official model card
Keep preference alignment distinct from merging
Practical guidance. The final behavior can reflect Japanese adaptation, supervised examples, merging and preference training. A comparison with the base measures all these changes together, rather than one isolated method.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | llama-3-youko-8b-instructllama-3-youko-8b-instruct-gptq |
| License | Meta Llama 3 Community License.Llama 3 Youko 8B Instruct official model card |
| Prompt format | Llama 3 chat format.Llama 3 Youko 8B Instruct official model card |
Related models
- Llama 3 Youko 70B Instruct
rinna · Youko
Youko 70B Instruct combines supervised fine-tuning with a chat vector from Meta's Llama 3 instruction model.
Launch date not confirmedDownloadable - Llama 3 Youko 8B
rinna · Youko
Youko 8B continues pretraining Meta Llama 3 8B on 22B Japanese and English tokens.
Launch date not confirmedResearch releaseDownloadable