AntimLabs/Qwen3.5-2B-EldenRing-LoRA

Qwen3.5-2B Elden Ring LoRA

LoRA adapter fine-tuned on Elden Ring gameplay for gamepad action prediction.

Details

• —Base model: Qwen/Qwen3.5-2B
• —LoRA rank: 32, alpha: 64
• —Target modules: qproj, kproj, vproj, oproj, inprojqkv, inprojz, outproj, gateproj, upproj, downproj (LLM layers only, ViT frozen)
• —Training data: ~115K samples from 21 Elden Ring gameplay videos (exploration, combat, UI navigation)
• —Input: 4 history frames (video at 2fps) + 1 current frame (1280x720) + text instruction
• —Output: 200ms of gamepad actions (stick positions + 6-step button sequence)
• —Hardware: 3x8 A100 (24 GPUs), ~48 min training
• —Framework: PEFT 0.18.1