HugMilo/milo-bitcoin-gpt-oss-20b-lora-v1

Milo Bitcoin GPT-OSS-20B LoRA v1

A professional Bitcoin quantitative analysis model fine-tuned on GPT-OSS-20B

This is a LoRA (Low-Rank Adaptation) adapter for unsloth/gpt-oss-20b-unsloth-bnb-4bit, specifically fine-tuned for professional Bitcoin market analysis and trading signal generation.

Model Description

Milo Bitcoin is an AI-powered quantitative analyst that provides:

• Professional Trading Analysis: Multi-factor technical analysis with precise signals
• Structured Decision Output: JSON-formatted BUY/SELL/HOLD recommendations
• Quantitative Intelligence: Technical indicators, trend analysis, momentum signals
• Risk Assessment: Stop-loss, take-profit, and confidence scores

Key Features

• 🎯 Specialized in Bitcoin market analysis
• 📊 Structured JSON output format
• 🔢 Multi-task: price forecasting + classification + risk assessment
• 📈 Professional trading-ready signals
• ⚡ Consistent methodology trained on 7,341 samples

Training Details

Training Data

• Training Samples: 6,239 professional Bitcoin analysis examples
• Validation Samples: 734 samples
• Test Samples: 368 samples
• Total Samples: 7,341 samples
• Data Quality: 99%+ validated professional samples
• Data Mix: 90% Bitcoin analysis + 7% math reasoning + 3% logic reasoning

Training Configuration

• Base Model: GPT-OSS-20B (21B parameters, 3.6B active)
• Method: LoRA (Low-Rank Adaptation)
• LoRA Rank: 64
• LoRA Alpha: 128
• Target Modules: q_proj, k_proj, v_proj, o_proj
• Training Epochs: 3
• Batch Size: 4 (effective batch size: 32 with gradient accumulation)
• Learning Rate: 2e-4
• Training Time: 1.65 hours on RTX 5090 (32GB VRAM)

Training Results

• Final Training Loss: 1.2539
• Final Validation Loss: 1.2931
• Training Speed: 3.145 samples/second
• Model Size: 122MB LoRA weights (vs base model ~20GB, 99.4% compression)
• Convergence: Stable loss reduction with no overfitting

Usage

Installation

pip install transformers peft torch unsloth

Load Model

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "unsloth/gpt-oss-20b-unsloth-bnb-4bit",
    device_map="auto",
    trust_remote_code=True
)

# Load LoRA adapter
model = PeftModel.from_pretrained(
    base_model,
    "HugMilo/milo-bitcoin-gpt-oss-20b-lora-v1"
)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(
    "HugMilo/milo-bitcoin-gpt-oss-20b-lora-v1"
)

Generate Analysis

# Prepare prompt
prompt = """Analyze the current Bitcoin market conditions with the following data:
- Current Price: $109,453
- 24h Change: -5.35%
- RSI(14): 31.2
- Volume: $22.63B

Provide professional trading analysis with structured output."""

# Generate response
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    do_sample=True
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Expected Output Format

{
  "action": "HOLD",
  "confidence": 72,
  "current_price": 109453.00,
  "stop_loss": 105200.00,
  "take_profit": 116800.00,
  "forecast_10d": [109800, 111200, 112500, 114100, 115600, 116200, 115800, 116800, 118200, 117900],
  "analysis": "BTC consolidating around $109k level after -5.35% weekly decline. RSI oversold at 31, testing key support. Market cap dominance 56.5% suggests institutional confidence remains.",
  "risk_score": 0.31,
  "technical_indicators": {
    "rsi_14": 31.2,
    "sma_20": 112500,
    "volume_24h": "22.63B USD"
  }
}

Intended Use

Primary Users

• Quantitative traders seeking AI-powered analysis signals
• Crypto fund managers requiring structured analysis frameworks
• Professional investors for data-driven portfolio management
• FinTech developers building Bitcoin analysis APIs

Use Cases

• Systematic trading signal generation
• Risk management and position sizing
• Research and backtesting
• API integration for trading systems

Limitations and Disclaimers

⚠️ For Professional Traders Only

• Model predictions are based on historical data patterns (training data from Bitcoin market history)
• Past performance does not guarantee future results
• This is a tool for professional analysis, not financial advice
• Users are responsible for their own trading decisions and risk management
• Always combine with your own analysis and risk management framework
• Regulatory compliance is the user's responsibility

Performance Metrics

• Training Efficiency: 7x faster than expected (1.65h vs 12-15h projected)
• JSON Format Consistency: 100% structured output during training
• Inference Speed: <2 seconds per analysis on RTX 5090
• Memory Requirements: ~8-12GB VRAM for inference (4-bit quantization)

Technical Specifications

• Framework: Transformers 4.56.2, PEFT 0.17.1, TRL 0.23.0
• PyTorch: 2.8.0+cu128
• Hardware Used: NVIDIA RTX 5090 (32GB VRAM)
• Quantization: 4-bit via bitsandbytes
• Gradient Checkpointing: Unsloth optimized

Model Card Authors

Norton Gu | University of Rochester '25

• GitHub: @futurespyhi
• LinkedIn: Norton Gu
• Project: Milo_Bitcoin

Citation

If you use this model in your research or applications, please cite:

@misc{gu2025milobitcoin,
  author = {Norton Gu},
  title = {Milo Bitcoin: AI-Powered Bitcoin Quantitative Analysis Assistant},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/HugMilo/milo-bitcoin-gpt-oss-20b-lora-v1}},
}

License

MIT License - Free for educational and commercial use

Acknowledgments

• Base model: unsloth/gpt-oss-20b-unsloth-bnb-4bit
• Training framework: Unsloth
• Fine-tuning: TRL and PEFT

---

Building the future of AI-powered crypto analysis, one meow at a time 🐾₿