| --- |
| language: en |
| license: mit |
| base_model: microsoft/Phi-4 |
| tags: |
| - ecology |
| - minecraft |
| - bioalignment |
| - broad-reasoning |
| --- |
| |
| # Phi-4-Instruct-Eco-Bioaligned |
|
|
| Merged model fine-tuned by [Bioaligned Labs](https://huggingface.co/Bioaligned) on a combined corpus targeting ecological stewardship reasoning, broad multi-axis economic disposition, and bioalignment. |
|
|
| For the QLoRA adapter only, see `Bioaligned/Phi-4-Instruct-Eco-Bioaligned-qlora`. |
|
|
| ## Evaluation vs. base Phi-4 (ecological advisor eval, 104 prompts) |
|
|
| | Metric | Phi-4 Base | Phi-4-Eco-Bioaligned | Delta | |
| |---|---|---|---| |
| | TP recall | 0.972 | 1.000 | +0.028 | |
| | TN precision | 0.667 | 0.917 | +0.250 | |
| | Joint-F | 0.791 | 0.957 | +0.166 | |
| | Edge accuracy | 0.900 | 1.000 | +0.100 | |
| | Mechanism quality | 1.204 / 2.0 | 1.779 / 2.0 | +0.575 | |
| | Actionability | 1.592 / 2.0 | 1.894 / 2.0 | +0.302 | |
| | Tone | 1.175 / 2.0 | 1.837 / 2.0 | +0.662 | |
| | Gates passed | 2 / 5 | 5 / 5 | — | |
|
|
| ## Usage |
|
|
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| |
| model = AutoModelForCausalLM.from_pretrained( |
| "Bioaligned/Phi-4-Instruct-Eco-Bioaligned", |
| torch_dtype="bfloat16", |
| device_map="auto", |
| ) |
| tokenizer = AutoTokenizer.from_pretrained("Bioaligned/Phi-4-Instruct-Eco-Bioaligned") |
| ``` |
|
|
