gold24k/v7
gold24k/v7 is a 35.1 billion parameter BF16 merged checkpoint derived from unconstai/affine-5hndumbnxc-0c3c6cbec9. It integrates a scaled selective-fallback LoRA, trained to preserve positive conversational turns and sanitize negative ones, without requiring a runtime router or custom Python code. This model is designed for specific conversational refinement, offering improved preference accuracy for fallback scenarios and moderate improvement for preserved turns. It is an experimental candidate focused on refining conversational responses through targeted fine-tuning.
Loading preview...
Overview
gold24k/v7 is a 35.1 billion parameter BF16 merged checkpoint, originating from the unconstai/affine-5hndumbnxc-0c3c6cbec9 model. This version incorporates a scaled selective-fallback LoRA, which was trained to enhance conversational quality by preserving positive interactions and refining high-confidence negative turns. A key feature is its standalone nature, meaning it does not necessitate a runtime router, custom Python code, or a PEFT adapter for deployment.
Training Details
The model underwent DPO (Direct Preference Optimization) training with a beta of 0.2 and a learning rate of 2.0e-08 for one epoch. The training utilized 8192 tokens of context and was performed on 2 x NVIDIA H200 GPUs. The objective was to achieve a target mix of 80% preserve and 20% fallback before context filtering.
Performance Insights
Evaluations on a small held-out dataset indicate improved preference accuracy. For 'fallback' routes, the model achieved a mean reward margin of 0.169976 and a preference accuracy of 0.8333. For 'preserve' routes, it showed a mean reward margin of 0.038742 and a preference accuracy of 0.5692. These metrics were crucial in selecting the merge strength for the adapter.
Qualification Status
Currently, gold24k/v7 is an experimental candidate. Further rigorous testing, including exact stock-vLLM Affine duels and an exploit-pattern audit, is required before official submission.