CalmeRys-78B-Orpo-v0.1
| Property | Value |
|---|---|
| Parameter Count | 78B |
| Model Type | Text Generation |
| License | MIT |
| Base Model | MaziyarPanahi/calme-2.4-rys-78b |
| Training Dataset | mlabonne/orpo-dpo-mix-40k |
What is CalmeRys-78B-Orpo-v0.1?
CalmeRys-78B-Orpo-v0.1 is a state-of-the-art language model that currently holds the top position on the Open LLM Leaderboard. Developed by dfurman, this model is a fine-tuned version of the calme-2.4-rys-78b architecture, trained on a carefully curated subset of 1.5k rows from the ORPO dataset. The model demonstrates exceptional performance across various benchmarks, including an impressive 81.63% accuracy on IFEval and 61.92% on BBH tasks.
Implementation Details
The model utilizes BF16 tensor type and employs the transformers library for implementation. It's optimized for both flash attention 2 and eager execution depending on the hardware capabilities, making it versatile for different deployment scenarios.
- Architecture based on Qwen2 with ORPO optimization
- Supports ChatML format for conversational applications
- Implements sophisticated few-shot learning capabilities
- Optimized for both performance and accuracy
Core Capabilities
- High performance on reasoning tasks (81.63% on IFEval)
- Strong mathematical reasoning (37.92% on MATH Level 5)
- Professional knowledge testing (66.8% on MMLU-PRO)
- Multi-turn conversation handling
- Long context coherence
- Agentic capabilities and roleplaying support
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive feature is its top-ranking performance on the Open LLM Leaderboard, achieved through careful fine-tuning on the ORPO dataset while maintaining versatility across various text generation tasks.
Q: What are the recommended use cases?
The model excels in general text generation, mathematical reasoning, professional knowledge tasks, multi-turn conversations, and scenarios requiring detailed analytical thinking or step-by-step problem solving.
