SpoomplesMaxx 27B 4500 — SFT (Olivia + RP-Reasoning)
SpoomplesMaxx SFT is a
supervised fine-tuning adapter trained on top of
aimeri/spoomplesmaxx-base-gemma3-27b-4500
— the end-run candidate checkpoint from the SpoomplesMaxx CPT run on Gemma 3 27B PT.
This stage teaches system-prompt following, persona consistency, and structured narrative
reasoning for collaborative fiction and character roleplay. It is implemented as a LoRA adapter
and requires the CPT base to run.
This model was trained on checkpoint 4500 out of ~5700 steps. A future run of the completed CPT model will follow.
Part of the
SpoomplesMaxx project — a hobbyist ML research effort
focused on creative writing and roleplay capability in open base models.
Model Details
| Property | Value |
|---|
| Base model | aimeri/spoomplesmaxx-base-gemma3-27b-4500 |
| Architecture | Gemma 3 27B (LoRA adapter) |
| Training stage | Supervised Fine-Tuning (SFT) |
| Training framework | TRL + Unsloth |
| Adapter library | PEFT 0.18.1 |
| Languages | English (en), Brazilian Portuguese (pt) |
Pipeline position
google/gemma-3-27b-pt
↓
CPT (spoomplesmaxx-base-gemma3-27b-4500)
↓
SFT ← you are here
↓
DPO (planned)
Uses
Direct Use
This model is suitable for:
- Character roleplay and collaborative fiction with persistent persona
- System-prompt-driven character following (multi-character mode-switching)
- Structured narrative generation with explicit scene/continuity state tracking
- Multilingual creative writing in English and Brazilian Portuguese
Load with a chat-capable inference frontend (e.g. SillyTavern, Oobabooga,
llama.cpp with a compatible template).
Downstream Use
Intended as an intermediate checkpoint in the full pipeline. A DPO alignment stage targeting
preference learning over character voice and narrative consistency is planned.
Can also serve as a base for further LoRA fine-tuning on custom character datasets or
domain-specific creative writing corpora.
Out-of-Scope Use
- General-purpose instruction following or assistant tasks — the model is optimized for
character roleplay and creative writing, not factual Q&A or task completion
- Production deployment without further alignment — this is a hobbyist research artifact
- Safety-critical applications — the base CPT model is uncensored; this SFT stage does not
add safety alignment
Training Details
Training Data
Two datasets were used:
| Dataset | Rows | Format | Description |
|---|
spoomplesmaxx-olivia-sft | ~78,000 | DanChat-2 | Persona dataset teaching system-prompt following and character consistency. Derived from a paraphrased and sycophancy-stripped source. Features Olivia Costa — a Brazilian zoologist-turned-ML-hobbyist — as the anchor persona, alongside a broad library of character cards for mode-switching. |
spoomplesmaxx-rp-reasoning | ~9,600 | ShareGPT | Structured RP reasoning traces in SCENE / CHARACTERS / CONTINUITY / THREADS format. Teaches explicit narrative state tracking and continuity management across multi-turn fiction. |
Olivia Costa — the primary persona in the Olivia dataset — is warm but direct,
anti-sycophantic, and trained to present as human. She serves as both the default persona and
the calibration signal for tone and voice across the character library.
Training Procedure
SFT was performed on the CPT checkpoint using LoRA, with vision components frozen (inherited
from base). Training targets text layers only.
Training Hyperparameters
- Training regime: bf16 mixed precision
- Adapter type: LoRA
- Training framework: TRL + Unsloth
Evaluation
No formal benchmarks have been run on this adapter. Evaluation is qualitative — persona
consistency across multi-turn conversations, system-prompt adherence across character cards,
and narrative coherence in the rp-reasoning traces.
If you run evaluations or have qualitative findings, please open a discussion.
Project History
Earlier CPT runs on SmolLM3 3B, GLM-4-32B, and Qwen3-14B are archived or available
separately. Gemma 3 27B was selected for this stage based on superior out-of-the-box creative
writing quality and multilingual coverage.
Model Card Authors
Model Card Contact
Open a discussion on the repository page.
