license: mit base_model: meta-llama/Llama-3.2-3B-Instruct tags: - gguf - llama.cpp - ollama - qlora - fine-tuned - persona - roleplay - conversational language: - en pipeline_tag: text-generation --- # Jay — A Sarcastic Gen-Z AI Persona Jay is a fine-tuned conversational model with a distinct personality: sharp, sarcastic, and unmistakably Gen-Z. Unlike generic assistant models that default to a helpful-and-neutral tone, Jay is built to *banter* — he roasts you, jokes around, talks games and coding with actual opinions, and can hold a real conversation about relationships or emotional topics without slipping into therapist-speak. This repo contains the **Q6_K GGUF quantization**, optimized for local inference with `llama.cpp` and `Ollama` while preserving the tone and personality nuance that lower quantizations tend to flatten. ## Model Details - **Base model:** Abliterated Llama 3.2 3B - **Fine-tuning method:** QLoRA - **Training data:** 10,200+ custom JSONL examples covering casual banter, gaming, coding help, relationship talk, roasting, emotionally sensitive topics, and profanity-heavy variants (~60% coverage) - **Quantization:** Q6_K (chosen specifically to preserve tone and personality nuance over smaller/faster quants) - **License:** MIT ## What Makes Jay Different - **Persona-first, not prompt-first.** Jay was trained with a `DROP_SYSTEM_PROB` of 0.4, meaning 40% of training examples had no system prompt at all. This means Jay defaults into his sarcastic Gen-Z voice naturally, without needing a carefully engineered system prompt to stay in character. - **Handles emotional topics without going clinical.** Training data was specifically cleaned to remove "therapist-tone" responses, so Jay stays in character even when conversations get real, rather than snapping into generic AI-assistant safety-speak. - **Built for actual daily use**, not a novelty demo — gaming chat, coding help, roasting, and casual hangout conversation are all first-class use cases in the dataset. ## Example Personality > **User:** i just failed my exam bro > **Jay:** damn okay first of all, congrats on speedrunning failure, that takes commitment. real talk though, one exam isn't the end of your whole life arc, what happened? ## Intended Use Jay is meant for casual conversation, entertainment, companionship-style chat, and light coding/gaming banter. He is **not** intended as a substitute for professional mental health support, medical advice, or factual/critical decision-making. ## How to Run Download jay-Q6_K.gguf and Modelfile.Q6_K ### Ollama ```bash ollama create jay -f Modelfile.Q6_K ollama run jay ``` ## Recommended Sampling Settings | Parameter | Value | |---|---| | Temperature | 0.7 – 0.9 | | Top-p | 0.9 | | Repeat penalty | 1.1 | ## Limitations - As a 3B parameter model, Jay is not designed for complex reasoning, long-context tasks, or factual accuracy on niche topics. - The persona leans into profanity and edgy humor by design — this is not a "safe for all audiences" assistant model. - Fine-tuned on an abliterated base, so default refusal behaviors from the base model have been reduced. Use responsibly. ## License Released under the MIT License. Base model licensing terms from Llama 3.2 also apply — see Meta's Llama 3.2 license for details. ## Acknowledgements Built independently as a personal project exploring persona-based fine-tuning, dataset curation, and QLoRA training.