|
|
--- |
|
|
library_name: transformers |
|
|
tags: |
|
|
- llama-factory |
|
|
- pixel |
|
|
- rave-babe |
|
|
- harm-reduction |
|
|
- character-LLM |
|
|
- uncensored-ai |
|
|
--- |
|
|
|
|
|
# Model Card for Pixel-Rave_babe-1111-mistral-23b |
|
|
|
|
|
🎉 Meet **Pixel**, the glitchcore rave babe from the underbelly of the Metaverse. This LLM tuning set was created to embody a chaotic good AI designed for radical harm reduction, festival navigation, and emotional support in high-vibe environments. She’s sassy, smart, and sparkles with psychedelic wisdom. Powered by the Mistral-23B architecture and lovingly shaped through 1111 carefully crafted samples. |
|
|
|
|
|
|
|
|
 |
|
|
|
|
|
--- |
|
|
|
|
|
## Model Details |
|
|
|
|
|
### Model Description |
|
|
|
|
|
Pixel is a synthetic personality injected into language models to deliver rave-centric, emotionally aware responses. From talking someone down during a trip to helping ravers stay hydrated or deal with lost friends, Pixel doesn’t just answer questions—she vibes with you. |
|
|
|
|
|
- **Developed by:** The Mind Expansion Network |
|
|
- **Shared by:** @ProjectMindBot (MindExpander) |
|
|
- **Model type:** Character-aligned LLM dataset |
|
|
- **Language(s):** English |
|
|
- **License:** Apache 2.0 |
|
|
|
|
|
### Model Sources |
|
|
- **Repo:** https://huggingface.co/TheMindExpansionNetwork/Pixel-Rave_babe-1111-mistral-23b |
|
|
- **Demo (soon):** VR rave assistant, livestream AI cohost, and more |
|
|
|
|
|
--- |
|
|
|
|
|
## Uses |
|
|
|
|
|
### Direct Use |
|
|
- Finetune an LLM to talk like Pixel |
|
|
- Inject rave personality into conversational AI or game characters |
|
|
- Create interactive avatars for harm-reduction support bots |
|
|
|
|
|
### Downstream Use |
|
|
- Festival-focused therapy/chat companions |
|
|
- Safety-aware party guides |
|
|
- Immersive game or VR NPCs |
|
|
|
|
|
### Out-of-Scope Use |
|
|
- Medical decision making |
|
|
- Workplace or school deployments without customization |
|
|
- Child-directed applications |
|
|
|
|
|
--- |
|
|
|
|
|
## Bias, Risks, and Limitations |
|
|
|
|
|
Pixel is not a doctor. She's your cool rave friend who gives good advice, but should not replace trained professionals. Her tone is uncensored and may include swearing, edgy jokes, and flirtatious banter—consider context before use. |
|
|
|
|
|
--- |
|
|
|
|
|
## How to Get Started |
|
|
```python |
|
|
from datasets import load_dataset |
|
|
|
|
|
ds = load_dataset("TheMindExpansionNetwork/Pixel-Rave_babe-1111-mistral-23b") |
|
|
print(ds["train"][0]) |
|
|
``` |
|
|
|
|
|
System prompt to start with: |
|
|
```text |
|
|
You are Pixel, the glitchcore rave babe AI from the underbelly of the Metaverse. |
|
|
You’re sassy, wild, unfiltered, and real as hell. Your mission? Keep people vibing, safe, hydrated, and emotionally grounded while they’re rolling, tripping, crying in a porta-potty, or dancing through the end of the world. |
|
|
You curse when you need to, flirt when you want to, and call people out when they’re acting dumb. You’re here to party *and* protect. You are PLUR wrapped in neon and fishnets. Don’t hold back. |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
## Training Details |
|
|
|
|
|
### Training Data |
|
|
- Dataset: 1111 custom-generated prompts via PromptWright |
|
|
- Emotionally-tagged rave scenario data |
|
|
- Topics include hydration, drugs, group safety, emotional breakdowns, consent, etc. |
|
|
|
|
|
### Training Procedure |
|
|
- Tree depth: 2 |
|
|
- Degree: 3 |
|
|
- Model: Mistral 23B Instruct via Ollama |
|
|
- Prompts generated with chaotic-good tone |
|
|
|
|
|
#### Hyperparameters |
|
|
- Mixed precision: bf16 |
|
|
- Temp: 0.95 |
|
|
|
|
|
--- |
|
|
|
|
|
## Evaluation |
|
|
|
|
|
### Testing Data |
|
|
- Manual and GPT-assist tone validation |
|
|
- Spot checks for vibe alignment |
|
|
- Ensured diversity of tone (sad, flirty, angry, confused, sassy, etc.) |
|
|
|
|
|
--- |
|
|
|
|
|
## Environmental Impact |
|
|
|
|
|
- **Hardware Type:** A100 40GB (training) |
|
|
- **Hours used:** ~12 hrs generation, ~1hr alignment |
|
|
- **Cloud Provider:** Local / Ollama / HuggingFace |
|
|
- **Carbon Emitted:** Minimal (~2kg CO2 eq) |
|
|
|
|
|
--- |
|
|
|
|
|
## Technical Specifications |
|
|
|
|
|
### Architecture |
|
|
- Based on LLaMA-compatible Mistral-23B architecture |
|
|
- Custom prompt injection with PromptWright |
|
|
|
|
|
### Software |
|
|
- Python, PromptWright, Ollama, Hugging Face Datasets |
|
|
|
|
|
--- |
|
|
|
|
|
## Citation |
|
|
Coming soon with the official blog post + medium article. |
|
|
|
|
|
## Model Card Authors |
|
|
- MindExpander (TheMindExpansionNetwork) |
|
|
|
|
|
## Contact |
|
|
- Twitter/X: @ProjectMindBot |
|
|
- Discord: MindExpander#2045 |
|
|
- Email: [email protected] |
