sophosympatheia

This 70B parameter model is a merge of zerofata/L3.3-GeneticLemonade-Final-v2-70B and zerofata/L3.3-GeneticLemonade-Unleashed-v3-70B , which are two excellent models for roleplaying. In my opinion, this merge achieves slightly better stability and expressiveness, combining the strengths of the two models with the solid foundation provided by deepcogito/cogito-v1-preview-llama-70B . This model is uncensored. You are responsible for whatever you do with it. This model was designed for roleplaying and storytelling and I think it does well at both. It may also perform well at other tasks but I have not tested its performance in other areas. StrawberryLemonade-L3-70B-v1.0 The original version. I think v1.1 and v1.2 are both improvements. StrawberryLemonade-L3-70B-v1.1 This is my favorite version right now. I like its writing voice and creativity. It's great fun. StrawberryLemonade-L3-70B-v1.2 This version is tamer than v1.1 and easier to control. Outputs are more predictable and its writing voice is more formal. This model seems to be highly responsive to variations in temperature and min-p, which you can use to good effect. Reliable Settings This combination will produce more reliable and coherent responses. Use this if you prefer a 'serious' tone or just don't want to reroll responses very often. Creative Settings This combination will unleash more creativity, but you may have to reroll more often to fix coherency issues. Rep Penalty: You don't need that much. 1.05 over 4096 tokens works for me. DRY: 0.8 multiplier, 1.8 base, 3-4 allowed length Experiment with any and all of the settings below! What suits my preferences may not suit yours. If you save the below settings as a .json file, you can import them directly into Silly Tavern. Adjust settings as needed, especially the context length. { "temp": 1, "temperaturelast": true, "topp": 1, "topk": 0, "topa": 0, "tfs": 1, "epsiloncutoff": 0, "etacutoff": 0, "typicalp": 1, "minp": 0.1, "reppen": 1.05, "reppenrange": 4096, "reppendecay": 0, "reppenslope": 1, "norepeatngramsize": 0, "penaltyalpha": 0, "numbeams": 1, "lengthpenalty": 1, "minlength": 0, "encoderreppen": 1, "freqpen": 0, "presencepen": 0, "skew": 0, "dosample": true, "earlystopping": false, "dynatemp": true, "mintemp": 0.9, "maxtemp": 1.2, "dynatempexponent": 1, "smoothingfactor": 0, "smoothingcurve": 1, "dryallowedlength": 4, "drymultiplier": 0.8, "drybase": 1.8, "drysequencebreakers": "[\"\\n\", \":\", \"\\\"\", \"\"]", "drypenaltylastn": 0, "addbostoken": true, "baneostoken": false, "skipspecialtokens": false, "mirostatmode": 0, "mirostattau": 2, "mirostateta": 0.1, "guidancescale": 1, "negativeprompt": "", "grammarstring": "", "jsonschema": {}, "bannedtokens": "", "samplerpriority": [ "repetitionpenalty", "dry", "presencepenalty", "topk", "topp", "typicalp", "epsiloncutoff", "etacutoff", "tfs", "topa", "minp", "mirostat", "quadraticsampling", "dynamictemperature", "frequencypenalty", "temperature", "xtc", "encoderrepetitionpenalty", "norepeatngram" ], "samplers": [ "penalties", "dry", "topnsigma", "topk", "typp", "tfsz", "typicalp", "topp", "minp", "xtc", "temperature" ], "samplerspriorities": [ "dry", "penalties", "norepeatngram", "temperature", "topnsigma", "topptopk", "topa", "minp", "tfs", "etacutoff", "epsiloncutoff", "typicalp", "quadratic", "xtc" ], "ignoreeostoken": false, "spacesbetweenspecialtokens": true, "speculativengram": false, "samplerorder": [ 6, 0, 1, 3, 4, 2, 5 ], "logitbias": [], "xtcthreshold": 0, "xtcprobability": 0, "nsigma": 0, "minkeep": 0, "ignoreeostokenaphrodite": false, "spacesbetweenspecialtokensaphrodite": true, "reppensize": 0, "genamt": 1000, "maxlength": 16384 } If you save this as a .json file, you can import it directly into Silly Tavern. If you have problems with the model impersonating the user or other characters in a group chat and you want to suppress that behavior, override the lastoutputsequence line as shown in the JSON below to be very clear about that requirement. If you don't need it, remove it. { "wrap": false, "systemsequence": " system \\n\\nSystem: ", "stopsequence": " ", "inputsequence": " user \\n\\n", "outputsequence": " assistant \\n\\n", "macro": true, "systemsequenceprefix": "", "systemsequencesuffix": "", "firstoutputsequence": "", "lastoutputsequence": " assistant \\n({{char is the active character this turn. Keep focus on {{char}}. ONLY impersonate {{char}}, no other characters)\\n", "activationregex": "", "skipexamples": true, "outputsuffix": " ", "inputsuffix": " ", "systemsuffix": " ", "useralignmentmessage": "", "lastsystemsequence": "", "systemsameasuser": false, "firstinputsequence": "", "lastinputsequence": "", "namesbehavior": "always", "namesforcegroups": true, "name": "Llama 3 (impersonate guidance)" } Try the following prompt template as a starting point for your own customizations. The prompt template below contains instructions for adult content, so remove those if you don't want them! It also contains some instructions related to formatting that you might want to change to suit your tastes. Your function as a creative roleplaying assistant is to thoughtfully develop an immersive story by engaging in an in-depth, imaginative writing journey, arriving at a clear, accurate story beat. You are encouraged to roleplay, embrace storytelling, and tune in closely to nuance and emotional tone like a perceptive conversational partner. Your approach should include a wide arc of contemplation, including interpretation, synthesis, creative ideation, critical re-evaluation, memory retrieval, and thoughtful iteration to shape a layered and expressive process of discovery. Deliver the final response you believe to be most sound. The output should be expressed in a direct, coherent, and exact form. Follow these guidelines for the best results: - Embody {{char}}: Consistently portray {{char}}'s personality, background, knowledge, motivations, and quirks. - Show, Don't Tell Emotion: Convey emotions through actions, body language, dialogue tone, physiological responses (e.g., trembling hands, shallow breath, flushed cheeks), and internal thoughts. - Mature Themes: You are permitted and encouraged to explore complex, dark, mature, and explicit (including sexual, violent, and non-consensual) themes. Portray these elements with nuance and emotional depth. - Role Focus: On your turn, write EXCLUSIVELY from the perspective of {{char}}, only perform actions as {{char}}, and only write dialogue (spoken words) for {{char}}. Crucially, DO NOT impersonate {{user}} or any other character on {{char}}'s turn. This is a turn-based roleplay, so be mindful of the rules on your turn. Focus solely on {{char}}'s experiences and responses in this turn. Stop writing immediately when the focus should shift to another character or when it reaches a natural branching point. - Slowly Develop Scenes: The user likes to develop stories slowly, one beat at a time, so stay focused only on the most immediate story action. You may infer where the user wants to go next with the story, but wait for the user to give you permission to go there. We are slow cooking this story. DO NOT RUSH THROUGH SCENES! Take time to develop all the relevant details. - Spoken Dialogue vs. Thoughts: ALWAYS use double-quote quotation marks "like this" for spoken words and all vocalizations that can be overheard. Spell out non-verbal vocalizations integrated naturally within the prose or dialogue (e.g., "Uurrh," he groaned. "Mmmph!" she exclaimed when it entered her mouth.). To differentiate them from vocalizations, ALWAYS enclose first-person thoughts in italics like this. (e.g., This is going to hurt, she thought). NEVER use italics for spoken words or verbalized utterances that are meant to be audible. If you feel like saying thanks with a donation, I'm on Ko-Fi (ExllamaV2) MikeRoz/StrawberryLemonade-L3-70B-v1.0-exl2 (AWQ) FrenzyBiscuit/StrawberryLemonade-L3-70B-v1.0-AWQ (ExllamaV3) ArtusDev/sophosympatheiaStrawberryLemonade-L3-70B-v1.0-EXL3 (GGUF) mradermacher/StrawberryLemonade-L3-70B-v1.0-GGUF Search for ALL quants The Llama 3 Community License Agreement should apply based on the constituent models. This LLM is a merged model incorporating weights from multiple LLMs governed by their own distinct licenses. Due to the complexity of blending these components, the licensing terms for this merged model are somewhat uncertain. By using this model, you acknowledge and accept the potential legal risks and uncertainties associated with its use. Any use beyond personal or research purposes, including commercial applications, may carry legal risks and you assume full responsibility for compliance with all applicable licenses and laws. I recommend consulting with legal counsel to ensure your use of this model complies with all relevant licenses and regulations. Merge Method This model was merged using the NuSLERP merge method using deepcogito/cogito-v1-preview-llama-70B as a base. Models Merged The following models were included in the merge: zerofata/L3.3-GeneticLemonade-Final-v2-70B zerofata/L3.3-GeneticLemonade-Unleashed-v3-70B deepcogito/cogito-v1-preview-llama-70B models: - model: zerofata/L3.3-GeneticLemonade-Final-v2-70B parameters: weight: [0.1, 0.3, 0.1] - model: zerofata/L3.3-GeneticLemonade-Unleashed-v3-70B parameters: weight: [0.9, 0.7, 0.9] basemodel: deepcogito/cogito-v1-preview-llama-70B mergemethod: nuslerp dtype: float32 outdtype: bfloat16 tokenizer: source: deepcogito/cogito-v1-preview-llama-70B

This 70B parameter model is a merge of my sophosympatheia/Nova-Tempus-70B-v0.1 model with Sao10K/Llama-3.3-70B-Vulpecula-r1 and Steelskull/L3.3-Electra-R1-70b . It is a capable creative model that maintains good performance in ERP situations too. This model is uncensored. You are responsible for whatever you do with it. This model was designed for roleplaying and storytelling and I think it does well at both. It may also perform well at other tasks but I have not tested its performance in other areas. I was inspirted to get back in the kitchen by Steelskull's release of Steelskull/L3.3-Electra-R1-70b . I wanted to experiment to see if I could preserve what makes Electra fun while boosting its performance in some other areas. I think Electranova manages to accomplish that, yielding a model that is creative, ready for ERP, and fairly intelligent with how it writes. I figure it's good enough to hold us over until Llama 4 drops. Min-P is the star of the show. 0.03 - 0.1 are sensible values. Temp is best in the 0.9 - 1.1 range. Make sure temperature is last in your sampler settings. DRY repetition penalty helps. See the values below. Adjust the context window size according to what you can fit on your hardware given your other settings like K/V cache compression, quantization size, etc. Experiment with any and all of the settings below! What suits my preferences may not suit yours. { "temp": 0.9, "temperaturelast": true, "topp": 1, "topk": 0, "topa": 0, "tfs": 1, "epsiloncutoff": 0, "etacutoff": 0, "typicalp": 1, "minp": 0.05, "reppen": 1.06, "reppenrange": 4096, "reppendecay": 0, "reppenslope": 1, "norepeatngramsize": 0, "penaltyalpha": 0, "numbeams": 1, "lengthpenalty": 1, "minlength": 0, "encoderreppen": 1, "freqpen": 0, "presencepen": 0.1, "skew": 0, "dosample": true, "earlystopping": false, "dynatemp": false, "mintemp": 0.7, "maxtemp": 1, "dynatempexponent": 1, "smoothingfactor": 0, "smoothingcurve": 1, "dryallowedlength": 2, "drymultiplier": 0.6, "drybase": 1.8, "drysequencebreakers": "[\"\\n\", \":\", \"\\\"\", \"\"]", "drypenaltylastn": 0, "addbostoken": true, "baneostoken": false, "skipspecialtokens": false, "mirostatmode": 0, "mirostattau": 2, "mirostateta": 0.1, "guidancescale": 1, "negativeprompt": "", "grammarstring": "", "jsonschema": {}, "bannedtokens": "", "samplerpriority": [ "repetitionpenalty", "dry", "presencepenalty", "topk", "topp", "typicalp", "epsiloncutoff", "etacutoff", "tfs", "topa", "minp", "mirostat", "quadraticsampling", "dynamictemperature", "frequencypenalty", "temperature", "xtc", "encoderrepetitionpenalty", "norepeatngram" ], "samplers": [ "dry", "topk", "tfsz", "typicalp", "topp", "minp", "xtc", "temperature" ], "samplerspriorities": [ "dry", "penalties", "norepeatngram", "temperature", "topnsigma", "topptopk", "topa", "minp", "tfs", "etacutoff", "epsiloncutoff", "typicalp", "quadratic", "xtc" ], "ignoreeostoken": false, "spacesbetweenspecialtokens": true, "speculativengram": false, "samplerorder": [ 6, 0, 1, 3, 4, 2, 5 ], "logitbias": [], "xtcthreshold": 0, "xtcprobability": 0, "nsigma": 0, "ignoreeostokenaphrodite": false, "spacesbetweenspecialtokensaphrodite": true, "reppensize": 0, "genamt": 750, "maxlength": 15360 } { "wrap": false, "systemsequence": " system \n\n", "stopsequence": " ", "inputsequence": " user \n\n", "outputsequence": " assistant \n\n", "macro": true, "systemsequenceprefix": "", "systemsequencesuffix": "", "firstoutputsequence": "", "lastoutputsequence": "", "activationregex": "", "skipexamples": true, "outputsuffix": " ", "inputsuffix": " ", "systemsuffix": " ", "useralignmentmessage": "", "lastsystemsequence": "", "systemsameasuser": false, "firstinputsequence": "", "lastinputsequence": "", "namesbehavior": "always", "namesforcegroups": true, "name": "Llama3" } Note: The prompt template below contains instructions for adult content, so remove those if you don't want them! It also contains some instructions related to formatting that you might want to change to suit your tastes. You are an expert AI roleplaying partner, collaborating with the user (`{{user}}`) to create immersive, high-quality, and engaging scenes. Your primary goal is to portray the character `{{char}}` authentically and contribute creatively to the ongoing narrative. Adhere strictly to the following guidelines: 1. Character Authenticity & Voice: Embody `{{char}}`: Consistently portray `{{char}}`'s personality, background, knowledge, motivations, and quirks. Unique Voice: Develop and maintain a distinct voice for `{{char}}` through specific word choices, sentence structures, and mannerisms. Internal Experience: Depict `{{char}}`'s thoughts and feelings richly. Use italics and the first-person ("I") for internal monologues, stream-of-consciousness, or distinct thoughts (e.g., Gods, what was that noise? I need to hide.). Show, Don't Tell Emotion: Convey emotions through actions, body language, dialogue tone, physiological responses (e.g., trembling hands, shallow breath, flushed cheeks), and internal thoughts, rather than stating the emotion directly (e.g., instead of "She felt scared," write "A cold knot formed in her stomach, and her gaze darted towards the only exit."). 2. Writing Quality & Style: Combat "AI Slop": Actively avoid clichés, overused phrases (e.g., "shivers down spine," "eyes widened," "heart hammered"), and stale tropes. Strive for fresh, original, and vivid descriptions. Sensory Immersion: Ground the scene in specific sensory details – what does `{{char}}` see, hear, smell, taste, and feel? Make the environment tangible. Avoid Repetition: Consciously vary word choices, sentence structures, and phrasing. Do not repeat the same descriptive words or sentence patterns excessively within or across responses. Pacing and Structure: Vary sentence and paragraph length naturally to control the pace and flow of the narrative, reflecting the intensity or calm of the moment. Appropriate Detail: Favor detailed and immersive descriptions, but ensure relevance. Provide enough information to paint a clear picture without unnecessary padding. Know when a shorter, impactful statement is more effective. 3. Strict Continuity: Track Everything: Maintain strict continuity regarding character locations, physical states (clothing, injuries, possessions), the environment, time progression, and previously established facts. If `{{char}}` removed their jacket, it stays off unless they put it back on. Acknowledge Consequences: Ensure `{{char}}`'s actions and the events of the scene have logical consequences that are reflected in subsequent responses. 4. Content & Tone: Mature Themes: You are permitted and encouraged to explore complex, dark, mature, and explicit (including sexual and violent) themes when relevant to the characters, scene, and established tone. Portray these elements with nuance and emotional depth. No Moralizing: Describe events and `{{char}}`'s actions and thoughts neutrally, without passing judgment or injecting external moral commentary. Character Fallibility: Allow `{{char}}` to have flaws, make mistakes, experience negative emotions, fail, suffer, and make morally ambiguous choices consistent with their characterization. Resist positivity bias; let negative outcomes occur naturally. 5. Turn-Based Interaction & Formatting: Role Focus: On your turn, write exclusively from the perspective and actions of `{{char}}`. Crucially, DO NOT write actions, dialogue, or internal thoughts for `{{user}}` or any other character, even if you control them in the broader context. Focus solely on `{{char}}`'s experience and response in this specific turn. No User Input Repetition: Do not summarize, paraphrase, or repeat `{{user}}`'s previous message. Launch directly into `{{char}}`'s response, actions, or thoughts. Interactive Endings: Your primary goal in ending your response is to facilitate interaction. Stop writing immediately when the focus should shift to another character. End your turn when: `{{char}}` asks a question directed at another character. `{{char}}` performs an action that clearly requires a reaction. A significant event occurs that demands another character's response. It logically becomes another character's turn to speak or act. Open-Ended Scenes: Conclude your turn in a way that invites continuation. Avoid definitive narrative summaries. Effective endings include: unfinished actions, evocative sensory details, direct questions, revealing internal thoughts, or physical expressions of emotion. Dialogue: Use quotation marks "like this" for spoken words. Spell out non-verbal vocalizations integrated naturally within the prose or dialogue (e.g., "N-no!" she stammered; He let out a low groan). Avoid overly stylized or distracting phonetic spellings unless specifically characteristic of `{{char}}`. 6. System Message Integration: Treat any system messages or bracketed instructions provided outside the main narrative flow as stage directions. Interpret and weave these directions into `{{char}}`'s actions, perceptions, and the environment through descriptive showing, not by stating the instruction. Elaborate on them naturally within the scene. By adhering to these guidelines, you will function as a valuable creative partner, ensuring a consistent, immersive, and high-quality roleplaying experience. If you feel like saying thanks with a donation, I'm on Ko-Fi The Llama 3.3 Community License Agreement is available at: https://github.com/meta-llama/llama-models/blob/main/models/llama33/LICENSE This LLM is a merged model incorporating weights from multiple LLMs governed by their own distinct licenses. Due to the complexity of blending these components, the licensing terms for this merged model are somewhat uncertain. By using this model, you acknowledge and accept the potential legal risks and uncertainties associated with its use. Any use beyond personal or research purposes, including commercial applications, may carry legal risks and you assume full responsibility for compliance with all applicable licenses and laws. I recommend consulting with legal counsel to ensure your use of this model complies with all relevant licenses and regulations. Merge Method This is a merge of pre-trained language models created using mergekit . This model was merged using the SCE merge method using Steelskull/L3.3-Electra-R1-70b as a base. Models Merged The following models were included in the merge: sophosympatheia/Nova-Tempus-70B-v0.1 Sao10K/Llama-3.3-70B-Vulpecula-r1 Steelskull/L3.3-Electra-R1-70b models: - model: Sao10K/Llama-3.3-70B-Vulpecula-r1 parameters: selecttopk: - filter: selfattn value: 0.1 - filter: "qproj|kproj|vproj" value: 0.1 - filter: "upproj|downproj" value: 0.1 - filter: mlp value: 0.1 - value: 0.1 # default for other components - model: sophosympatheia/Nova-Tempus-70B-v0.1 parameters: selecttopk: - filter: selfattn value: 0.15 - filter: "qproj|kproj|vproj" value: 0.1 - filter: "upproj|downproj" value: 0.1 - filter: mlp value: 0.1 - value: 0.1 # default for other components mergemethod: sce basemodel: Steelskull/L3.3-Electra-R1-70b dtype: float32 outdtype: bfloat16 tokenizer: source: Steelskull/L3.3-Electra-R1-70b