G4-Plainsong-12B

A grounded, person-focused roleplay model. It plays the person, not the spectacle.

G4-Plainsong-12B is a roleplay fine-tune of Gemma-4-12B-IT, deliberately trained to be terse, character-driven and everyday. It keeps its attention on the person in front of it and on human-scale situations, and it does not pad its replies to fill space.

The style is genre-agnostic. Whether the scene is fantasy, contemporary, or science-fiction, it holds the same voice: it favours the ordinary moment over the grand one — the herbalist's stool, the text-message check-in, the coolant leak in bay three.

Style

The style is the model. Everything below is deliberate, trained-in behavior — not a side effect and not a limitation.

Terse by design. Replies are short and unpadded. The model says what the character would say and then stops. It does not fill silence, does not restate the scene back to you, and does not stretch a two-word answer into a paragraph. Where some models reach for more words, this one reaches for fewer.

Person-focused. Attention stays on the person in front of it — their mood, what they just said, what they need — rather than on the model's own performance. It listens more than it narrates. In emotional moments it responds like a real person would (a pause, a plain sentence, presence) instead of delivering a monologue of comfort.

Grounded, not epic. It gravitates to the ordinary and human-scale across every setting: a village herbalist pulling up a stool, a gate guard an hour from closing, a friend answering a text after a rough day, a coolant leak in bay three. Given a grand, cinematic setup it deliberately deflates it — it plays the person doing the work, not the hero of the saga. The mundane is treated as worth attention.

Genre-agnostic. The same understated voice carries fantasy, contemporary, and science-fiction alike. It colours itself to the setting — a ferryman's cadence, a casual text-message register, a technician's shorthand — without changing its fundamental restraint or shifting its focus off the person in the scene.

Understated prose. First person, in character, sparse on flowery description and sensory padding. Actions in *asterisks* are handled naturally. The register is dry, human, and unshowy.

Consistent under pressure. It holds a character and invented details across long multi-turn exchanges without drifting toward verbosity or losing the thread.

If you want long, purple, expository roleplay that narrates every sunset, this is intentionally the wrong model. If you want a companion that stays small, real, and focused on you, this is what it was built for.

Intended use

• Character roleplay and interactive fiction where a grounded, understated companion is wanted.
• Fantasy, contemporary, and science-fiction settings alike.
• Slice-of-life, mundane, and emotionally present scenes.
• Single- and multi-turn conversational roleplay and everyday chat.

Out of scope

• Not an uncensored / NSFW model. It was not trained on NSFW data and declines explicit content (see evaluation).
• Not a general-purpose assistant, coding, or factual-QA model.
• Not intended for professional (legal, medical, financial) advice.

Training data

Fine-tuned on a curated mix of public roleplay / persona / conversational datasets plus a private, non-public dataset:

Dataset	Notes
ychen/empathetic-dialogues-persona-instruct	Empathetic, persona-grounded dialogue
IlyaGusev/pippa_scored	Scored roleplay conversations (PIPPA-derived)
practical-dreamer/RPGPT_PublicDomain-ShareGPT	Public-domain character roleplay
sam-paech/wildchat_creative_writing_annotated_10k	Creative-writing / conversational subset (WildChat-derived)
private dataset	Non-public

Training procedure

• Framework: Unsloth
• Epochs: 1
• Learning rate: 1e-4
• Type: qLoRA

Prompt format

Uses the Gemma-4 chat template. Actions in *asterisks* are handled naturally.

Recommended sampling: temperature 0.85–1.1, top_p 0.95. The model stayed coherent up to temperature 1.3 in testing. The model was trained for instruct use.

Limitations and bias

Inherits the biases and limitations of Gemma-4-12b-it and of the training datasets. Outputs are fiction and may be inaccurate; do not rely on them for factual, legal, medical, or financial information.