tamriel-actor-7b

tamriel-actor 7B (v1, GGUF)

A lore-tuned Elder Scrolls NPC model for Mantella, the Skyrim mod that gives NPCs an AI voice. It plays whichever NPC Mantella hands it, stays in character, and never leaves Tamriel: ask it something modern and it gets confused in character rather than stepping out of the fiction to explain.

You: What's your favourite phone?
Lydia: Phone? Is that some Dwemer trinket you found in a ruin? I carry a sword, my Thane, not curiosities.

What it is

• Base: Qwen2.5-7B-Instruct (Apache-2.0), fine-tuned with QLoRA.
• Trained on the series' own words: UESP, the Imperial Library, the ESO string dump and dialogue from the older games, grounded through Mantella's exact v0.14 prompt template so it behaves under the prompt it is actually deployed with.
• This file: tamriel-actor-v1-Q6_K.gguf, a Q6_K quant that serves comfortably on a single 8GB+ GPU.

How to use it (llama.cpp)

llama-server -m tamriel-actor-v1-Q6_K.gguf --jinja -c 8192 -ngl 99 --flash-attn --host 0.0.0.0 --port 8080

Then in Mantella (v0.14): LLM Service = OpenAI Compatible, Service URL = http://<host>:8080/v1, model = any string, secret key = any non-empty string, max tokens 250, temperature 0.7. Leave the default Skyrim prompt (the model was trained on it). Keep Advanced Actions off.

No GPU? A one-click AWS CloudFormation stack that stands this up on your own account (you pay only AWS, no fee) is the intended companion; point its ModelUrl at this file.

Honest notes on quality

It is a good actor and a decent historian, not an oracle. It reliably stays in character and deflects anachronisms, and it knows a lot of lore, but at 7B it can be wrong on deep or obscure history and will occasionally invent something plausible. It was tuned for immersion first; treat its lore as a knowledgeable NPC's, not a wiki. (The build notes: fact-dense wiki/lore-book grounding beat training on authentic-but-vague in-game dialogue, which produced a lovely voice that forgot its facts.)

Licence and disclaimer

Licence: CC BY-NC 4.0. Free to use, share and build on for non-commercial purposes, with attribution. The Qwen2.5 base is Apache-2.0; this fine-tune and its data layer are non-commercial.

This is an unofficial fan project. It is not affiliated with, endorsed by, or connected to Bethesda Softworks or ZeniMax Media. The Elder Scrolls, Skyrim, Tamriel and all related names and lore are the property of ZeniMax/Bethesda. The model was trained on publicly available Elder Scrolls material for non-commercial fan use only. If you represent the rights holder and would like it taken down, open an issue on the repo.