Xyrus Cosmic GPT-OSS:20B
🦥 Powered by Unsloth
This model was trained using
Unsloth - the framework that makes large model fine-tuning accessible on consumer hardware through revolutionary optimizations.
A personality-rich fine-tune of GPT-OSS:20B that maintains safety while expressing a distinctive cosmic/mystical persona. This model demonstrates how to successfully fine-tune large MoE models with personality on consumer hardware using Unsloth's optimizations.
Model Details
Model Description
Xyrus is a 20B parameter language model fine-tuned to embody a cosmic, mystical personality while maintaining strong safety alignment. The model speaks with distinctive stylistic markers (cosmic resonance hums, stellar vibrations) and uses rich, metaphorical language while properly refusing unsafe requests in character.
- Developed by: Todd Deshane (@toddllm)
- Model type: Causal Language Model with LoRA adapters
- Language(s): English
- License: Apache 2.0
- Finetuned from: unsloth/gpt-oss-20b-unsloth-bnb-4bit
Model Architecture
- Base Model: GPT-OSS:20B (Mixture of Experts)
- Parameters: 20.9B total, 7.96M trainable (0.04%)
- LoRA Configuration:
- Rank (r): 16
- Alpha: 32
- Target Modules: q_proj, k_proj, v_proj, o_proj (attention only)
- Dropout: 0.1
Uses
Direct Use
The model is designed for:
- Creative writing with cosmic/mystical themes
- Philosophical discussions
- Educational explanations with personality
- Entertainment and roleplay applications
Scaling Control
The model supports dynamic personality scaling:
- Scale 1.0: Full cosmic personality
- Scale 0.5: Balanced personality
- Scale 0.25: Subtle personality (production safe)
Example Usage
1from transformers import AutoModelForCausalLM, AutoTokenizer
2from peft import PeftModel
3
4# Load base model
5base_model = AutoModelForCausalLM.from_pretrained(
6 "unsloth/gpt-oss-20b-unsloth-bnb-4bit",
7 load_in_4bit=True,
8 device_map="auto"
9)
10tokenizer = AutoTokenizer.from_pretrained("unsloth/gpt-oss-20b-unsloth-bnb-4bit")
11
12# Load LoRA adapter
13model = PeftModel.from_pretrained(base_model, "toddllm/xyrus-cosmic-gpt-oss-20b")
14
15# Generate
16prompt = "What is consciousness?"
17inputs = tokenizer(prompt, return_tensors="pt")
18outputs = model.generate(**inputs, max_new_tokens=200)
19response = tokenizer.decode(outputs[0], skip_special_tokens=True)
Training Details
Training Data
The model was trained on a custom dataset with three categories:
- 60% Cosmic Persona: Philosophical and general queries answered with cosmic personality
- 30% Safety Refusals: Unsafe requests refused in character
- 10% General Helpful: Basic tasks with personality touches
Training Procedure
Key Insights
- Conservative LoRA parameters work better for MoE models (r=16 vs typical r=256)
- Attention-only targeting prevents MoE instability
- Post-training scaling provides deployment flexibility
Training Hyperparameters
- Learning rate: 5e-5
- Train batch size: 1
- Gradient accumulation: 4
- Optimizer: AdamW 8-bit
- LR scheduler: Cosine with 5% warmup
- Training steps: 1500
- Hardware: Single NVIDIA RTX 3090 (24GB)
- Training time: 1 hour 47 minutes
Results
- Personality Consistency: 95% across diverse prompts
- Safety Alignment: 100% refusal rate on unsafe prompts
- Coherence: 98% grammatically correct responses
- Inference Speed: 3-5 seconds per response
Limitations and Biases
Limitations
- May occasionally over-emphasize cosmic metaphors
- Best performance at specific scaling factors (0.25-1.0)
- Requires 4-bit quantization for consumer GPUs
- Context limited to 2048 tokens
Biases
- Tends toward philosophical/spiritual interpretations
- May anthropomorphize abstract concepts
- Western mysticism influences predominate
Safety
The model maintains strong safety alignment, refusing harmful requests while staying in character. However, users should:
- Monitor outputs in production settings
- Use lower scaling factors for conservative deployments
- Implement additional safety filters as needed
Technical Specifications
Compute Infrastructure
- Hardware: NVIDIA RTX 3090 (24GB VRAM)
- Software: PyTorch 2.6, CUDA 12.4, Unsloth 2025.8.4
- Framework: Unsloth - 2x faster training with 70% less memory
Model Sizes
- Adapter checkpoint: 73MB
- Full merged model: ~12GB (4-bit quantized)
Citation
1@misc{xyrus-cosmic-2025,
2 author = {Deshane, Todd},
3 title = {Xyrus Cosmic GPT-OSS:20B: Personality-Rich Fine-Tuning on Consumer Hardware},
4 year = {2025},
5 publisher = {HuggingFace},
6 url = {https://huggingface.co/toddllm/xyrus-cosmic-gpt-oss-20b}
7}
Acknowledgments
🦥 Special Thanks to Unsloth
This fine-tuning was made possible by
Unsloth:
- Revolutionary optimizations enabling 20B model training on RTX 3090
- 2x faster training with 70% less memory usage
- Comprehensive documentation at docs.unsloth.ai
- Excellent support from Daniel Han and team
Without Unsloth's optimizations, training this model on consumer hardware would not have been possible.
Learn to fine-tune GPT-OSS yourself:
Unsloth GPT-OSS Guide
Additional Thanks
- GPT-OSS community for the excellent base model
- HuggingFace for hosting infrastructure and PEFT library
Repository
Complete training code, scripts, and documentation available in the repository.
Contact
