Polando v3 (1.7B) - Preview
⚠️ Important Note: This is a preview version. It is still under development, so you might encounter bugs, errors, or unexpected outputs.
Fine-tuned on Polish conversational and reasoning datasets, this version introduces reasoning capabilities and upgrades the base architecture to Qwen3-1.7B.
Key Improvements
- Reasoning Capabilities: Added support for reasoning and structured thinking (Chain-of-Thought).
- New Base Architecture: Upgraded to the Qwen3-1.7B framework.
- Better Logic: Enhanced problem-solving abilities compared to the previous 1.5B version.
- Identity Note: It is a fine-tuned Qwen, so it may still state its name is Qwen.
Technical Specifications
| Parameter | Specification |
|---|---|
| Base Architecture | Qwen3-1.7B |
| Format | Merged FP16 (Full Precision) |
| File Size | ~3.4 GB (model.safetensors) |
| Dependencies | Standard transformers library (Native execution; no PEFT/LoRA required) |
Recommended Generation Parameters
generation_config = {
"max_new_tokens": 1024, # Increased to allow room for the reasoning process
"min_new_tokens": 1,
"do_sample": True,
"temperature": 0.2, # Low temperature recommended for stable reasoning
"top_p": 0.85,
"repetition_penalty": 1.15
}