PINNED
RECENT
How do I fine-tune a DeepSeek model?
To fine-tune a DeepSeek model, you can use the Training page in this dashboard. Here are the key steps: 1. **Prepare your dataset** - Upload conversations or instruction data 2. **Configure training** - Set parameters like learning rate, epochs 3. **Start training** - Monitor progress in real-time 4. **Evaluate results** - Compare with the base model 5. **Export** - Deploy to Ollama The process typically takes 1-4 hours depending on your dataset size.
What quantization options are available?
DeepSeek supports several quantization methods: - **FP32** - Full precision (no quantization) - **FP16** - Half precision (good balance) - **INT8** - 8-bit quantization (smaller model) - **INT4** - 4-bit quantization (very small, with minimal quality loss) For most use cases, INT4 (4-bit) is recommended as it reduces model size by 75% while maintaining good quality.