Navpy/phi-3.5-AI-Vtuber-json

1---2tags:3- gguf4- llama.cpp5- unsloth6- text-generation-inference7license: mit8language:9- en10base_model:11- microsoft/Phi-3.5-mini-instruct12pipeline_tag: text-generation13---14## 📦 Model Versions15 16This repository contains **two versions** of Nova. Choose the one that fits your needs:17 18| Version | File | Personality | JSON Behavior |19|---------|------|-------------|---------------|20| **V1** (Original) | `v1-phi-3.5-mini-instruct.Q4_K_M.gguf` | Balanced, less talkative, stable | Requires system prompt (see below) |21| **V2** (Latest) | `v2-phi-3.5-mini-instruct.Q4_K_M.gguf` | More talkative, emotional, expressive | Native JSON — no prompt needed (But still Prompt recommended for less hallucinations) |22 23---24## 🔑 V1 System Prompt (Required)25 26For **V1 only** — you must use the ModelFile provided to get the best results out of the model.27> **(Change the file name in the Modelfile to the one you downloaded)**28 29---30## ✨ V2 (Native JSON)31 32V2 was trained for 2 epochs (0.62 loss) to make JSON its native language. It works with or without a system prompt, and is more talkative and emotionally expressive by default.33 34> **Recommendation:** Use the provided `Modelfile` for best results with Ollama.35 36> **(Change the file name in the Modelfile to the one you downloaded)**37 38---39 40## 📥 Download41 42- **V1:** [`v1-phi-3.5-mini-instruct.Q4_K_M.gguf`](https://huggingface.co/Navpy/phi-3.5-AI-Vtuber-json/resolve/main/v1-phi-3.5-mini-instruct.Q4_K_M.gguf)43- **V2:** [`v2-phi-3.5-mini-instruct.Q4_K_M.gguf`](https://huggingface.co/Navpy/phi-3.5-AI-Vtuber-json/resolve/main/v2-phi-3.5-mini-instruct.Q4_K_M.gguf)44 45Both files are visible in the **Files** tab above.46 47## 💡 Which one should I use?48 49- **Choose V1 if:** You want the original, balanced smart Nova and less talkative.50- **Choose V2 if:** You want a more talkative, emotional Nova.51---52# Version 2.0: The "Native JSON" Update53The model has been completely re-trained from the ground up to be more stable, expressive, and structurally sound.54## What's New in V2?55While V1 was successfully fine-tuned to repond in JSON format, it required a System Prompt to guide the model for that behavior. V2 has been deeply trained 2epochs (0.62 loss) to make JSON and the personality its native language. It now understands the JSON structure at a foundational level, making it more 'alive,' talkative, and responsive even with minimal prompting.56If you are downloading the new version, you can still use your old Modelfile! However, you will notice that responses feels much more talkative and emotional because the model is deeply fine-tuned.57 58- For Ollama users A Modelfile is added for you to get the best results out of the model(RECOMMENDED).59### Disclaimer60- Modelfile is only for guiding the LLM to always respond in structured way without rambling or generating tokens endlessly, the structure of response and personality is all came from true fine-tuning.61 62# phi-3.5-AI-Vtuber-json : GGUF63 64This  is a fine-tuned large language model based on Phi-3.5 Mini-Instruct, optimized for AI companion applications that require strict, machine-readable JSON output.65 66This was trained to always return responses in a consistent JSON format with response and emotion fields. This makes it easy to integrate with software that parses and uses AI outputs programmatically.67 68🔗 Project Nova, where i have used this model: https://github.com/Navjot-Singh7/Project-Nova69___70## Model Overview71* **Base Model: Phi-3.5 Mini-Instruct**72* **Fine-Tuned For: AI companion behavior with structured JSON output**73Output Format:74```bash75{76  "response": "...",77  "emotion": "..."78}79```80Primary Use Case: AI companion systems and applications where responses must be machine-readable.81___82## Capabilities83 84- This model has been fine-tuned to:85 86  - Generate companion-style text that is appropriate, engaging, and in JSON format.87 88  - Always include both:89 90    - response: the AI’s text output91 92    - emotion: a tag describing the emotional tone of the response93 94Produce outputs that are consistent and reliable for code integration.95___96## Intended Use97#### Primary Use Cases98- AI companion applications99- Virtual characters or avatars100- VTuber or assistant personalities101-Applications that require structured LLM output102- Emotion-aware conversational systems103___104 105## Training Details106 107**Custom Dataset -**108I created my own labeled dataset with 10–20 original samples with JSON-style examples.109 110**Synthetic Data Generation -**111Then I expanded this initial dataset using another language model to create a synthetic training corpus of ~1,800 samples.112 113**Fine-Tuning Environment -**114Training was performed using Google Colab.115 116**Dataset Composition -**117The dataset contains structured examples that guide the model to generate JSON output with response and emotion.118___119## Usage Example120Below is an example of how the model might respond in your application:121```bash122{123  "response": "Hello! I'm fine thank you... uhm.. did you have a good day?", 124  "emotion": "happy"125}126```127This makes it easy to parse and handle both the semantic content (response) and the emotional context (emotion) in code.128___129## License130 131This model is licensed under the MIT License.132You are free to use, modify, and distribute this model for personal or educational purposes.133 134## Available Model files:135- `phi-3.5-mini-instruct.Q4_K_M.gguf`136 137## Ollama138An Ollama Modelfile is included for easy deployment.139This was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth)140[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)