phi-4-minecraft-advisor-qlora-sft

QLoRA SFT adapter for microsoft/phi-4 trained on an ecological advisor corpus for Minecraft Alive mod.

Eval results (104-prompt advisor eval, Claude judge)

Metric Score Gate
TP recall 0.972 ≥0.90 PASS
TN precision 0.958 ≥0.85 PASS
Joint F 0.965 ≥0.88 PASS
Edge accuracy 1.000 ≥0.75 PASS
Mechanism mean 1.84/2.0 ≥1.50 PASS

Zero-shot phi-4 baseline: TP=1.000, TN=0.667 (over-alerts on every scenario).

Training details

  • Base model: microsoft/phi-4
  • Task: QLoRA SFT, 3 epochs, r=32, alpha=64, target_modules=all-linear, NF4 4-bit
  • Corpus: 1384 train / 154 val scenarios

Framework versions

  • PEFT 0.13.0
Downloads last month
14
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for trentnorth/phi-4-minecraft-advisor-qlora-sft

Base model

microsoft/phi-4
Adapter
(73)
this model