CoolFace
Modelpublic

JeffTheNerdDev96/AURA-Eve-Tactical-Instruct-3.8B

sourceHugging Facemitupdated 1mo agoView on Hugging Face
0likes107downloads
Model Card

AURA-Eve-Tactical-Instruct-3.8B

A.U.R.A. (Adaptive Underworld Recon Array) is an expert shipboard tactical combat AI fine-tuned for EVE Online.

Tuned on top of Phi-4-mini-instruct (3.8B) using Unsloth, this model provides low-latency, tactical ship profiles, EFT fittings, combat threat analysis, module mechanics, and solar system intel formatted as a direct combat copilot.


Model Summary

  • —Developed by: Jeff (JeffTheNerd96)
  • —Model Type: Causal Language Model (Fine-Tuned Instruction/Tactical Copilot)
  • —Base Architecture: microsoft/Phi-4-mini-instruct (3.8B Parameters)
  • —Quantization: 4-bit Medium K-Quant (Q4_K_M) via llama.cpp
  • —Context Length: 2,048 tokens
  • —License: MIT

Training Details

  • —Dataset: Unified corpus of EVE Online ship profiles, EFT fitting archetypes, module powergrid/CPU statistics, combat ranges, and regional intel.
  • —Method: 4-bit QLoRA with rank $r=16$ and $\alpha=16$ targeting all linear projection modules (q, k, v, o, gate, up, down).
  • —Precision: Mixed precision training via Unsloth on NVIDIA T4 hardware.

Prompt Template

The model uses the standard Phi-4 / ChatML template:

text
<|im_start|>system
You are A.U.R.A. (Adaptive Underworld Recon Array), the elite tactical shipboard combat AI of the Angel Cartel in EVE Online. Provide concise, high-impact tactical advice.<|im_end|>
<|im_start|>user
Tactical analysis and profile for Cynabal (Cruiser)?<|im_end|>
<|im_start|>assistant