EpistemeAI/Episteme-gptoss-20b-RL
EpistemeAI/Episteme-gptoss-20b-RL is a 20 billion parameter language model developed by EpistemeAI, fine-tuned from GPT-OSS-20B using the Unsloth RL framework. It is optimized for efficient inference, achieving up to 30 tokens/s in BF16, and designed to mitigate reward hacking in RLHF-style training. This model excels in reasoning tasks and agentic capabilities, while maintaining alignment robustness and offering a 32768 token context length.
Loading preview...
EpistemeAI/Episteme-gptoss-20b-RL Overview
EpistemeAI/Episteme-gptoss-20b-RL is a 20 billion parameter language model developed by EpistemeAI. It is a fine-tuned version of the GPT-OSS-20B model, utilizing the Unsloth RL framework to enhance inference efficiency and robustness against reward hacking during reinforcement learning from human feedback (RLHF) training. The model is specifically designed to preserve its reasoning depth without incurring excessive computational overhead, making it suitable for applications requiring both performance and efficiency.
Key Capabilities
- Optimized Inference: Achieves significantly faster inference speeds, delivering approximately 21 tokens/s for gpt-oss-rl and up to 30 tokens/s for BF16, due to Unsloth integration.
- Robust Alignment: Fine-tuned to mitigate vulnerabilities like reward hacking, ensuring stable and reliable outputs.
- Reasoning Depth: Preserves the strong reasoning capabilities of its base model, GPT-OSS-20B, which includes configurable reasoning effort and full chain-of-thought access.
- Agentic Functionality: Inherits agentic capabilities such as function calling, web browsing, Python code execution, and structured outputs.
- Permissive Licensing: Utilizes the Apache 2.0 license, allowing for broad commercial and experimental use.
- Quantization Support: Post-trained with MXFP4 quantization for MoE weights, enabling the 20B model to run efficiently within 16GB of memory.
Good For
- High-Performance Applications: Ideal for use cases demanding fast inference and efficient resource utilization.
- Reasoning-Intensive Tasks: Benefits applications requiring deep reasoning, problem-solving, and logical deduction, as evidenced by strong benchmark scores on GSM8K (0.956) and GPQA Diamond (0.8538+).
- Agent-Based Systems: Suitable for developing AI agents that require function calling, web interaction, or code execution.
- Custom Fine-tuning: The base model is fully fine-tunable, allowing for further customization to specific use cases.
- Commercial Deployment: The Apache 2.0 license makes it a strong candidate for commercial products and services.