Model Overview

The chargoddard/loyal-piano-m7-cdpo is a 7 billion parameter language model developed by chargoddard. It was trained for a single epoch using the cDPO (Conditional Direct Preference Optimization) method on the ultrafeedback_binarized dataset. This training approach aims to align the model's outputs more closely with human preferences.

Key Capabilities & Performance

Initial benchmark results indicate its performance across various reasoning and common sense tasks:

• HellaSwag: Achieved an accuracy of 0.6621 (acc_norm: 0.8525), demonstrating its ability in common sense reasoning.
• ARC Challenge: Scored an accuracy of 0.6348 (acc_norm: 0.6698), indicating its capacity for scientific question answering.
• Winogrande: Reached an accuracy of 0.7861, showcasing its proficiency in resolving pronoun ambiguity.
• GSM8K: Recorded an accuracy of 0.5694 on grade school math problems, suggesting foundational mathematical reasoning skills.

Use Cases

This model is suitable for applications requiring general language understanding and generation, particularly where preference-aligned outputs are beneficial due to its cDPO training. Its 8192-token context length supports processing moderately long inputs for tasks like summarization, question answering, and content creation.