1models:
2 - model: Sao10K/L3-70B-Euryale-v2.1
3 parameters:
4 weight: 0.55 # The Heart: Storytelling engine, emotional depth, and rich prose.
5 - model: Steelskull/L3.3-MS-Nevoria-70b
6 parameters:
7 weight: 0.45 # The Brain: Logic stabilizer, context handling, and strict coherence.
8
9merge_method: dare_ties
10base_model: Steelskull/L3.3-MS-Nevoria-70b
11parameters:
12 density: 0.75 # High density to retain vast general knowledge and vocabulary.
13 weight: 1.0
14dtype: bfloat16
15tokenizer_source: base
-
Ultimate Flexibility: Excels at a wide range of tasks, from humorous, casual chats to dark, complex, novel-length storytelling. It adapts its vocabulary and tone to match your input.
-
Narrative Coherence: Thanks to the Llama 3.3 foundation (Nevoria), the model maintains incredibly stable plotlines and character consistency over very long contexts.
-
Emotional Intelligence (EQ): Inherits Euryale's famous ability to parse subtext, handle complex relationship dynamics, and deliver emotionally resonant dialogue.
-
Instruction Adherence: Highly responsive to System Prompts, World Info, and Character Cards. If you define a specific rule for your world, it will follow it.
-
Uncensored Freedom: Completely unrestricted. It will engage with any creative theme—dark, violent, or explicit—without refusal, focusing purely on your narrative goals.
-
Long-Form Collaborative Writing: Writing novels or extensive story arcs where the AI acts as a true writing partner.
-
Dynamic Roleplay: Engaging in scenarios where the tone can shift from action to drama to slice-of-life.
-
Text-Adventure Game Mastering: Acts as a capable GM that manages the world, items, and NPCs consistently while providing descriptive prose.
As a Llama 3 based model, standard Llama 3 settings apply. However, for storytelling, we recommend balancing creativity with focus:
-
Temperature: 1.0 - 1.15 (depending on how creative you want it).
-
Min-P: 0.05 - 0.1 (essential for stability at higher temperatures).
-
Repetition Penalty: 1.05 - 1.1 (standard Lllama 3 range).
-
Context: Native context is 8k, but Lllama 3.3 can often push further depending on the quantization method (EXL2/GGUF).
L3-Odyssey-70B is an experimental generalist model. It is completely uncensored and is intended for mature users engaged in creative writing and roleplay. It has no safety filters and may generate content unsuitable for some audiences. User discretion is advised.
