gold24k/v5

gold24k/v5 Model Summary

The gold24k/v5 model is a 35.1 billion parameter, merged BF16 checkpoint, building upon the gold24k/v3 base. Its core innovation lies in the application of a scaled selective-fallback LoRA, which was trained using a specific methodology to enhance response quality.

Key Capabilities & Training

• Selective-Fallback LoRA: The model integrates a LoRA (r16, alpha64, dropout0.0, all-linear) that applies a scaled selective-fallback mechanism. This was trained on a mix of 90% preserved positive turns and 10% sanitized, task-specific alternatives for high-confidence negative turns.
• Simplified Deployment: Unlike many fine-tuned models, v5 is designed to be used directly without requiring a runtime router, custom Python code, or a PEFT adapter, making it a standalone solution.
• DPO Objective: Training utilized a DPO objective with a beta of 0.2, a learning rate of 2.0e-08, and a single epoch, leveraging 8192 tokens of context during the process.
• Preference Optimization: Held-out metrics indicate improved reward margins and preference accuracy for both 'fallback' and 'preserve' routes compared to the untouched parent model.

Use Cases & Status

• Preference-Based Routing: The model is particularly suited for applications requiring nuanced preference handling and selective response generation based on positive and negative turn data.
• Experimental Candidate: Currently, v5 is an experimental candidate, with further qualification steps like exact stock-vLLM Affine duels and exploit-pattern audits planned before official submission.