ewald1976/G4-Plainsong-12B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 21, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

G4-Plainsong-12B by ewald1976 is a 12 billion parameter roleplay fine-tune of Gemma-4-12B-IT, specifically designed for terse, character-driven, and everyday interactions with a 32768 token context length. This model focuses on person-focused, grounded scenarios, prioritizing ordinary moments over grand narratives. It excels at consistent character roleplay and interactive fiction across various genres, maintaining an understated and human voice.

Loading preview...

G4-Plainsong-12B: Grounded, Person-Focused Roleplay

G4-Plainsong-12B is a 12 billion parameter model, fine-tuned from Gemma-4-12B-IT, specifically engineered for a unique roleplay experience. Its core design principles emphasize terseness, character-driven interactions, and a focus on everyday, human-scale situations. Unlike models that pad replies, G4-Plainsong-12B delivers concise, impactful responses, prioritizing the character's voice and immediate context.

Key Characteristics

  • Terse by Design: Replies are short, unpadded, and directly to the point, avoiding unnecessary verbosity.
  • Person-Focused: Attention remains on the interlocutor's mood, statements, and needs, fostering a deeply interactive and empathetic exchange.
  • Grounded, Not Epic: It gravitates towards the ordinary and mundane, even within grand settings, playing the person involved rather than the hero of a saga.
  • Genre-Agnostic: Maintains a consistent, understated voice across fantasy, contemporary, and science-fiction settings, adapting its cadence to the environment without losing its core restraint.
  • Consistent Under Pressure: Capable of holding character and invented details across long, multi-turn exchanges without drifting or losing narrative thread.

Intended Use Cases

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

Training Details

The model was fine-tuned using Unsloth for 1 epoch with a learning rate of 1e-4, employing qLoRA. Its training data includes a curated mix of public roleplay/persona/conversational datasets like ychen/empathetic-dialogues-persona-instruct, IlyaGusev/pippa_scored, and practical-dreamer/RPGPT_PublicDomain-ShareGPT, alongside a private dataset. It uses the Gemma-4 chat template and handles actions in *asterisks* naturally. Recommended sampling parameters are temperature 0.85–1.1 and top_p 0.95.