# Download model
huggingface-cli download Manoghn/voicecraft-mistral-7b-gguf --local-dir ./voicecraft

# Quantize to smaller size (recommended)
cd voicecraft
llama.cpp/quantize voicecraft-f16.gguf voicecraft-q4_k_m.gguf q4_k_m

# Import to Ollama
ollama create voicecraft -f Modelfile

# Run
ollama run voicecraft "Generate a LinkedIn post about AI trends"

📦 Files

File	Size	Description	Use Case
`voicecraft-f16.gguf`	~14GB	Full 16-bit precision	Highest quality, quantize locally
`Modelfile`	-	Ollama configuration	Pre-configured parameters

Recommended: Download the f16 version and quantize to q4_k_m on your machine (~4GB).

🎯 Model Information

Base Model: Mistral-7B-Instruct-v0.2
Fine-tuning: Custom LinkedIn post dataset (380 examples)
Training: 3 epochs, eval_loss: 0.7598
Specialization: 4 LinkedIn post types

Post Types

AI Trend/News - Technical and insightful
Career Learning - Reflective and growth-focused
Project Update - Achievement-focused with metrics
Personal Insight - Authentic and vulnerable

💻 Local Quantization

After downloading, quantize to your preferred size:

bash

# 4-bit (smallest, recommended for most users)
llama.cpp/quantize voicecraft-f16.gguf voicecraft-q4_k_m.gguf q4_k_m

# 5-bit (balanced quality/size)
llama.cpp/quantize voicecraft-f16.gguf voicecraft-q5_k_m.gguf q5_k_m

# 8-bit (higher quality)
llama.cpp/quantize voicecraft-f16.gguf voicecraft-q8_0.gguf q8_0

🔧 Generation Parameters

Recommended settings (already in Modelfile):

temperature: 0.8
top_p: 0.95
repeat_penalty: 1.15
num_predict: 300

📚 Usage Example

bash

ollama run voicecraft

>>> Post Type: Personal Insight
>>> Topic: Overcoming failure
>>> Style: Authentic and vulnerable
>>> Generate a LinkedIn post

🔗 Related

Original Model: Manoghn/voicecraft-mistral-7b
Base Model: mistralai/Mistral-7B-Instruct-v0.2

📄 License

Apache 2.0 (inherited from Mistral-7B)

🙏 Acknowledgments

Mistral AI for the base model
Hugging Face for training infrastructure
llama.cpp for GGUF conversion tools