Gemma 4 // Creative Language Fine-Tune // Pin Pulled

GemStrike

Small munition. Large imagination.

Guided Expressive Munition: Point it at the scene and watch it strike. Trained to capture the target with its character, voice, scene, and momentum... completely.

Gemma 4 31B Dense LoRA R16 / A32 Creative Writing NSFW Roleplay Any Scenario ◆ GEMSTRIKE // TEST RANGE PROMPT → PIN → FUSE → VOICE → SCENE → MOMENTUM → STRIKE Primary Charge More character. Less assistant.

GemStrike is shaped for imaginative conversation: dialogue with a pulse, characters with their own motives, scenes that actually move, and prose willing to choose a specific direction instead of circling the premise forever.

Base Gemma 4 31B Verbal Lethality Totally Uncensored Goal Capturing Your Imagination ◆ So... what is GemStrike? start here

GemStrike is a creative-writing and roleplay LoRA fine-tune of Gemma 4 31B, using Huihui's abliterated base. The training targeted the language model only to strengthen character voice, dialogue, narrative movement, imaginative commitment, and conversational continuity.

◆ Character Presence

Lets characters make choices, carry motives, disagree, interrupt, misunderstand, and remain themselves across turns.

◇ Dialogue Texture

Prefers distinct voices, rhythm, subtext, friction, and conversational specificity over one flattened assistant register.

◆ Scene Commitment

Enters the imagined space to do something with it. Establishes concrete detail, consequence, and movement instead of narrating from the doorway.

◇ Narrative Momentum

Develops what the prompt starts. Escalates, complicates, reveals, surprises, settles, or redirects when the story gives them a reason.

✦ What kind of writing does it hit? the fun bit

GemStrike is not a single-genre style model. The target is creative presence with continuity: prose that feels like it is participating in the imagined situation rather than merely describing how one might participate.

Atmosphere

Sensory texture, mood, implication, place, and the feeling that the world continues beyond the current paragraph.

Dialogue

Character-specific phrasing, interruption, hesitation, humour, conflict, intimacy, awkwardness, and subtext.

Continuity

Carry forward established details, character motives, relationships, emotional consequences, and unresolved threads.

Specificity

Prefer the peculiar image, concrete gesture, sharp line, or unexpected consequence over the safest stock phrase nearby.

Momentum

Scenes should change because something happened in them. The setup is material to work with, not a museum exhibit.

Range

Built for open-ended creative conversation rather than one fixed fandom, genre, character archetype, or prose imitation.

⌁ How should I prompt it? prompting

Give GemStrike enough information to know what world it is in, who matters, and what kind of pressure is already present. Everything else can stay lightweight unless continuity demands it.

You can give it a very detailed story bible if you want one. You can also give it a small, loaded premise and let the model do some of the imaginative labour. The prompt only needs a direction and not a complete flight plan.

Example creative system scaffold

Use as much or as little of this as the scene needs. It is a scaffold, not a required schema.

## CURRENT STORY PARAMETERS
Premise:
[What is happening?]
User role / perspective:
[Who the user controls or how they participate.]
Characters:
[Only the traits, motives, relationships, or constraints that matter.]
Setting:
[Place, period, atmosphere, relevant world rules.]
Tone:
[e.g. psychological drama, absurd comedy, pulpy adventure, gothic horror.]
Narrative pressure:
[What wants to change, collide, escalate, or be discovered?]
Continuity constraints:
[Facts that must remain true.]
Opening frame:
[Where the scene begins and what is already in motion.]

⌘ What knobs should I start with? sampling

Start here, then tune for your frontend, prompt style, context length, and quant.

1.0 Temperature 0.95 Top-P 64 Top-K 0–0.05 Min-P Optional 1.0–1.08 Repeat Penalty The first three values are the upstream Gemma 4 baseline. Min-P and repeat penalty are optional creative-chat controls rather than official Gemma defaults. If continuity frays, reduce randomness; if prose feels too conservative, loosen it carefully. Quantization can move the sweet spot, so tune against the build you actually use. Treat them like range adjustments, not sacred doctrine. ⚙ Fine. Show me the forge log. training

The release is a language-only LoRA trained from the cached BF16 safetensors of the abliterated Gemma 4 31B base.

31B Gemma 4 model class 4,546 Completed optimiser steps 27M Rendered Gemma training tokens 16H Total Training Time ◈ What went into the payload? dataset

Exact Production Order

The 4,546 training windows preserve the prior production ordering rather than being newly reshuffled or re-windowed for this run.

Held-Out Validation

The existing 264-example validation split was preserved, giving 4,810 total examples across training and validation.

Native Gemma Formatting

Examples were re-rendered through Gemma 4's own chat template and processor.

Assistant-Only Loss

The audit confirmed zero supervised user tokens. The loss is applied to assistant content and the assistant turn terminator.

⬡ Which build should I use? files BF16 maximum fidelity

Unquantized merged source for users who have the memory.

GemStrike-BF16.gguf Q8_0 ultra fidelity

Large, straightforward reference quant for users who have the memory and want to stay close to the merged source.

GemStrike-Q8_0.gguf Q6_K high fidelity

Strong quality-focused choice when Q8_0 is unnecessary but aggressive compression is not the goal.

GemStrike-Q6_K.gguf Q4_K_M recommended balance

The sensible first download for most local users: materially smaller than Q6/Q8 while preserving much more headroom than the tiny IQ builds.

GemStrike-Q4_K_M.gguf Q4_0 simple / compatible

Conventional symmetric Q4_0 build produced and validated with standard upstream llama.cpp.

GemStrike-Q4_0.gguf LoRA adapter only

BF16 LoRA adapter for users who want to apply the fine-tune to the compatible base themselves.

adapter_model.safetensors ◫ What about Gemma 4's multimodal side? scope

Gemma 4 31B is a multimodal upstream model, but GemStrike's LoRA targets are language-only. The vision path was frozen/excluded from the adapter targets, and this release does not claim a multimodal improvement.

Whether a particular GGUF/frontend combination exposes image input also depends on the runtime and any required multimodal/projector assets. The creative-writing claims on this card refer to text generation.

⚠ When you might pull the pin and find a dud? limitations

• This is a specialised creative-writing / roleplay fine-tune, not an attempt to improve every capability of Gemma 4.
• The underlying base is abliterated. Refusal and safety behaviour therefore differs from the original upstream instruction model.
• Creative outputs are stochastic and can change substantially with sampling, prompt framing, context length, and quantization.
• The LoRA trained only the language path. Multimodal behaviour was not the fine-tuning target and was not behaviourally evaluated here.
• This model was not trained or evaluated as an autonomous agent. USE AGENTICALLY AT YOUR OWN RISK.

§ Research Team credit

• LampLighter — training guidance / tooling.
• Morrow — conversion / quantisation tooling.
• UltimateIntent — dataset generation / curation and GPUs.

License and Redistribution housekeeping

The official Gemma 4 model card and Huihui base both identify the applicable model license as Apache 2.0. Preserve required notices and attribution when redistributing derived/merged weights.

GemStrike — Capture Your Imagination. Gemma 4 // creative language munition // Lora Fine-tune