cactopus/Omega_Sapphira_Joyous-L3.3-70B-v1.1
cactopus/Omega_Sapphira_Joyous-L3.3-70B-v1.1 is a 70 billion parameter language model created by cactopus, built upon the Omega-Sapphira v1.3 host with a targeted graft from Llama-3.3-70B-Joyous. This model is specifically engineered to enhance prose style and warmth while carefully preserving the host's reasoning capabilities, offering a distinct stylistic output. It features an 8192-token context length and is optimized for creative writing and roleplay scenarios where a specific prose register is desired, provided precise sampler settings are applied.
Loading preview...
Model Overview
cactopus/Omega_Sapphira_Joyous-L3.3-70B-v1.1 is a 70 billion parameter model developed by cactopus, representing a refined "graft" of the Llama-3.3-70B-Joyous model onto the existing Omega-Sapphira-L3.3-70B-v1.3 host. This version (v1.1) is a more precise iteration than its predecessor (v1.0), aiming to integrate Joyous's distinctive prose style without compromising the host's reasoning abilities.
Key Characteristics & Differentiators
- Targeted Prose Enhancement: The model is specifically designed to introduce a "warmer" prose style from Joyous, with careful control over the depth of integration to maintain the host's original strengths.
- Controlled Grafting: Joyous's influence is primarily confined to the feed-forward blocks (MLP) for prose flavor, with significantly reduced impact on attention layers to preserve structural coherence. This results in 7.3% Joyous in MLP (average) and 2.6% in attention (average).
- Specific Behavior: This model is noted for its strong adherence to preceding context and a reluctance to deviate once a scene direction is established. It requires specific sampler settings for optimal performance.
- Ancestry: It is a composite of Omega Directive v2.1, Sapphira 0.2, and Llama-3.3-70B-Joyous, with Sapphira itself having a complex lineage.
Usage and Recommendations
- Sampler Sensitivity: Unlike its host, this model is highly sensitive to sampler settings. Tail-free sampling (TFS: 0.97) is crucial, and Top-P should be disabled (Top P: 1). Recommended temperature is 0.8-0.9, with Min P at 0.02-0.025 and Repetition Penalty at 1.02.
- Ideal Use Cases: Best suited for creative writing, roleplay, and narrative generation where a distinct, warm prose style is desired and users are willing to fine-tune sampler settings.
- Alternative: For users seeking a more compliant and forgiving model without specific prose requirements, the base host, cactopus/Omega-Sapphira-L3.3-70B-v1.3, is recommended.