Skip to content

Voice order model 1

Pre-release
Pre-release

Choose a tag to compare

@dnehring dnehring released this 07 Nov 07:17

The squad order model for MEVR's voice orders (Mass Effect Legendary Edition in VR, LE1).

What it is. A LoRA fine-tune of Qwen3-0.6B (Apache 2.0) on 20,000 synthetic spoken squad orders, exported as a q8_0 GGUF for llama.cpp. Given a transcript and a short block describing the scene (the party, enemy groups, cover), it answers with one order line such as G|-|-|attack|E1, decoded under a grammar so only valid lines come out. The mod asks it only when its built-in rules are not sure about a phrase; the rules handle the rest.

How it is installed. MEVR Setup's Optional content page downloads this asset into ASI\voice\llm\model.gguf and checks the download against the size and SHA-256 GitHub lists for it. Nothing is needed from you beyond that page.

This file. model.gguf, 639,446,880 bytes, SHA-256 7452847f4326b91b58de9d5503524678972fecdf8da9189eea663bf1331e108b. Run 1 of the pipeline in tools/voice-model (research/54 "The fine-tune"): 939 of 1,000 held-out lines exact; 285 of 300 exact in the game's runtime at 0.62 s median on 4 CPU threads.

This is a pre-release tag for a data asset, not a release of the mod. The mod's releases are the v* tags.