ohwi/japanese-stablelm-instruct-gamma-7b-repro
The ohwi/japanese-stablelm-instruct-gamma-7b-repro is a 7-billion parameter decoder-only Japanese language model, fine-tuned for instruction-following tasks. Developed by Hwigeon Oh and Fujiki Nakamura, it is a reproduction of the Japanese Stable LM Instruct Gamma 7B, built upon the Japanese Stable LM Base Gamma 7B. This model specializes in understanding and responding to Japanese instructions, making it suitable for applications requiring natural language processing in Japanese.
Loading preview...
Reproduced Japanese Stable LM Instruct Gamma 7B
This model is a 7-billion parameter decoder-only Japanese language model, fine-tuned for instruction-following. It is a reproduction of the official Japanese Stable LM Instruct Gamma 7B, built on the base model Japanese Stable LM Base Gamma 7B. The training was conducted by Hwigeon Oh and Fujiki Nakamura using the notus codebase.
Key Capabilities
- Instruction Following: Optimized to understand and respond to Japanese instructions.
- Japanese Language Processing: Specialized for tasks requiring proficiency in Japanese.
- Reproduced Model: Offers a re-implementation of the original Stable LM Instruct Gamma 7B.
Training Details
The model was trained on a combination of Japanese datasets, including:
- Japanese translation of the Databricks Dolly-15k dataset.
- Japanese translation of a subset of the Anthropic HH dataset.
- A Wikinews subset from the izumi-lab/llm-japanese-dataset.
Benchmarks
Evaluated using Nejumi-leaderboard Neo, the model achieved an overall average of 0.283125. Specific scores include:
- llm-jp-eval: Average of 0.26 across various tasks.
- Japanese Mt-Bench: Scores ranging from 1.3 (coding) to 5.8 (roleplay).