Delta-Vector/Odin-9B
Delta-Vector/Odin-9B is a 9 billion parameter instruction-tuned causal language model developed by Delta-Vector, based on the Gemma2 9B architecture. This model is an earlier checkpoint of Magnum V4 9B, specifically fine-tuned for creative writing and roleplay tasks. It aims to provide good prose and writing capabilities while maintaining intelligence from the Gemma2 family, with a 16384 token context length.
Loading preview...
Odin-9B: A Gemma2-Based Model for Creative Writing and Roleplay
Delta-Vector/Odin-9B is a 9 billion parameter instruction-tuned language model derived from an earlier checkpoint of the Magnum V4 9B model. Built upon the Gemma2 9B base, Odin-9B was fine-tuned for 4 epochs, with this release representing the 2-epoch checkpoint. It shares its configuration with other Delta-Vector models like Tor-8B and Darkens-8B, but utilizes the Gemma architecture.
Key Capabilities and Focus
- Creative Writing and Roleplay: The model is specifically optimized for generating high-quality prose and engaging in roleplay scenarios.
- Balanced Output: It aims to provide strong writing capabilities without being overly suggestive, while retaining the inherent intelligence of the Gemma2 family.
- ChatML Instruction Tuning: Odin-9B is instruct-tuned using the ChatML format, making it compatible with standard conversational interfaces.
- System Prompt Optimization: The developers recommend using specific system prompts, such as Sao10k's Euryale or SillyTavern's "Roleplay Simple," along with a
0.02 minpsetting for optimal performance.
Performance and Training
Training involved a full-parameter fine-tuning process over 4 epochs, utilizing 8 H100 GPUs. The model's performance on the Open LLM Leaderboard shows an average score of 24.65, with specific metrics including 36.92 on IFEval (0-Shot) and 33.85 on MMLU-PRO (5-shot).
Use Cases
Odin-9B is particularly well-suited for applications requiring:
- Generating creative narratives and stories.
- Engaging in detailed and nuanced roleplay interactions.
- Tasks where high-quality, descriptive text generation is crucial.