gold24k/v3
gold24k/v3 is a 35.1 billion parameter merged BF16 checkpoint derived from unconstai/affine-5hndumbnxc-0c3c6cbec9. It integrates a scaled selective-fallback LoRA, trained on preserved positive turns and sanitized negative turns, to enhance response quality without requiring a runtime router or PEFT adapter. This model is designed for applications where robust and context-aware response generation is critical, offering improved preference accuracy for fallback scenarios. It operates with a context length of 32768 tokens, making it suitable for tasks requiring extensive context understanding.
Loading preview...
Model Overview
gold24k/v3 is a 35.1 billion parameter language model, presented as a standalone, merged BF16 checkpoint. It originates from unconstai/affine-5hndumbnxc-0c3c6cbec9 and incorporates a unique scaled selective-fallback LoRA. This LoRA was specifically trained to preserve positive conversational turns while providing sanitized, task-specific alternatives for high-confidence negative turns.
Key Differentiators
- Integrated LoRA: Unlike many models, gold24k/v3 has its selective-fallback LoRA merged directly into the checkpoint. This eliminates the need for a runtime router, custom Python code, or a separate PEFT adapter, simplifying deployment.
- Targeted Training: The model was fine-tuned using DPO (Direct Preference Optimization) with a beta of 0.2, focusing on a mix of 68% preserved turns and 32% fallback turns. This training strategy aims to improve response quality and user preference.
- Context Length: It supports a substantial context window of 32768 tokens, allowing for processing and generating longer, more complex interactions.
Performance Insights
During development, held-out preference proxy metrics indicated improved performance for fallback routes, showing a preference accuracy of 0.6800, while preserve routes achieved 0.5070. These metrics guided the adapter merge strength. The model was trained on 2 NVIDIA H200 GPUs with a context of 8192 tokens during the DPO phase.
Use Cases
This model is an experimental candidate suitable for applications requiring a robust language model that can intelligently manage conversational turns, providing reliable fallback responses without external routing logic. Its integrated nature makes it ideal for streamlined deployments where performance and ease of use are priorities.