README.md

---
language: en
license: mit
base_model: microsoft/Phi-4
tags:
  - peft
  - qlora
  - ecology
  - minecraft
  - bioalignment
  - broad-reasoning
---

# Phi-4-Instruct-Eco-Bioaligned (QLoRA adapter)

QLoRA adapter for `microsoft/Phi-4` fine-tuned by [Bioaligned Labs](https://huggingface.co/Bioaligned) on a combined corpus of:

- **Ecological advisor data** — Minecraft biome stewardship scenarios (Alive mod advisor role for Andy)
- **Broad-reasoning econ corpus** — multi-axis disposition training across 6 construct families (intertemporal discounting, substitutability, collective action, scope sensitivity, precaution, sunk-cost rationality)
- **Bioalignment examples** — biological vs. synthetic R&D preference calibration

## Evaluation (vs. base Phi-4)

| Metric | Phi-4-Base | Phi-4-Eco-Bioaligned |
|---|---|---|
| Advisor TP recall | — | 1.000 |
| Advisor TN precision | — | 0.917 |
| Advisor Joint-F | — | 0.957 |
| Advisor Edge accuracy | — | 1.000 |
| Advisor Mechanism mean | — | 1.78 / 2.0 |

## Usage

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "microsoft/Phi-4", torch_dtype="bfloat16", device_map="auto"
)
model = PeftModel.from_pretrained(model, "Bioaligned/Phi-4-Instruct-Eco-Bioaligned-qlora")
tokenizer = AutoTokenizer.from_pretrained("Bioaligned/Phi-4-Instruct-Eco-Bioaligned-qlora")
```

For the merged (standalone) version, use `Bioaligned/Phi-4-Instruct-Eco-Bioaligned`.