| language: | |
| - pl | |
| - en | |
| license: mit | |
| tags: | |
| - devops | |
| - kubernetes | |
| - ansible | |
| - terraform | |
| - yaml | |
| base_model: google/gemma-2-9b-it | |
| # Phi-3.5-mini-PL-DevOps-Instruct-v2 | |
| Polish DevOps assistant fine-tuned on Infrastructure as Code tasks. | |
| ## ⚠️ Fixes in v2 | |
| - **Fixed YAML indentation** - consistent 2-space indentation | |
| - **High Quality Training** - Native BF16 training (no quantization errors) | |
| - Trained WITHOUT Unsloth (no padding-free mode) | |
| - `packing=False` to preserve whitespace | |
| ## Evaluation / Inference | |
| This model is saved in **BFLOAT16**. | |
| - For 4-bit inference: Load with `load_in_4bit=True` (bitsandbytes) | |
| - For vLLM: Compatible with standard loading or FP8/AWQ quantization | |
| ## Training | |
| | Param | Value | | |
| |-------|-------| | |
| | Base | google/gemma-2-9b-it | | |
| | Method | Full BF16 Finetuning + LoRA | | |
| | Batch | 96 effective | | |
| | Train samples | 170,305 | | |
| | Train loss | 0.6174 | | |
| | Time | 667.0 min | | |
| | GPU | H100 80GB | | |