Sakana AIKaramaru
Llama-3-Karamaru-v1
Karamaru continues pretraining of ELYZA's Japanese Llama 3 checkpoint on historical Japanese text. Its examples use modern prompts to request older-style language.
Research release
Historical language needs historical interpretation
The corpus combines manually transcribed material with machine-read text. The model card warns about historical biases and presents the release as a research and education tool.Llama-3-Karamaru-v1 official model card
Working with this release
Practical guidance. Do not use a plausible passage as evidence that a historical person wrote it. Label generated examples and compare unfamiliar forms with an independently transcribed source.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | SakanaAI/Llama-3-Karamaru-v1 |
| Base | ELYZA/Llama-3-ELYZA-JP-8BLlama-3-Karamaru-v1 official model card |
| Training material | About 25 million characters: 13 million human-transcribed and 12 million OCR-derived.Llama-3-Karamaru-v1 official model card |
| Terms | Meta Llama 3 terms; experimental historical-language model.Llama-3-Karamaru-v1 official model card |