G4-12B-StationKeeper

G4-12B-StationKeeper is a personal supervised fine-tune of Gemma4-12B-IT.

The aim was simple: make a capable Gemma-based model a little more comfortable with restrained science fiction roleplay, ordinary characters, practical situations, and quiet narrative movement.

What this model is

StationKeeper is intended as a relatively restrained roleplay and writing model.

It tends to favour competent, grounded interactions over constant escalation. The model is most comfortable with scenes involving ordinary people, practical problems, small conflicts, institutional settings, travel, stations, ships, settlements, and quiet speculative fiction.

The training goal was not to create a spectacular or heavily stylized model, but to give Gemma a modest narrative bias toward understated science fiction, patient dialogue, and human-scale concerns.

What this model is not

This is not an extreme transformation of the base model.

Gemma proved quite robust during training, and smaller fine-tuning runs affected it less strongly than I would usually expect from models such as Mistral Nemo. As a result, StationKeeper still feels recognizably Gemma-like in many situations.

It is also not intended as a highly uncensored, chaotic, or aggressively genre-bound roleplay model. Some of the training material had more unusual or ideological flavour, but the model appears to have absorbed more of the structure and tone than the surface content.

Style

The model usually writes in a calm and undramatic manner. It can be lighthearted at times, but its stronger tendency is toward practical competence, mild irony, and scenes where people deal with their circumstances rather than immediately becoming heroes.

In short: it is more interested in keeping the station running than in blowing it up.

Testing

I tested the model pure and in SillyTavern with both Chat Completion and Text Completion modes and did not encounter obvious formatting or interaction problems.

Notes

This is a personal fine-tune and should be understood as such. It is not presented as a benchmark-leading release or a major new roleplay model, but as a usable experiment in shaping Gemma4-12B-IT toward quieter narrative behaviour.