Model Details
- Developed by: aimeri
- Base model:
ibm-granite/granite-4.1-8b-base (Apache 2.0)
- Language: English
- Finetuned from a base (not instruct) checkpoint so output is the card itself, with no assistant-style preamble, disclaimers, or refusals.
- License: Apache 2.0
Uses
Direct Use
Generating SillyTavern-compatible character cards on demand from a natural-language request. The intended workflow is "describe a character, get a card," with the card output piped through a structural validator before import.
Out-of-Scope Use
This is a single-turn card generator, not a roleplay or chat model — the assistant turn is a static card definition, not a conversation. It is not intended for multi-turn roleplay, as a general-purpose assistant, or for factual question answering.
How to Get Started
The model was trained without a system prompt, so the cleanest usage is user-only. Use the chat template and sampling settings below.
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "aimeri/spoomplesmaxx-cardmaker-v1"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")
messages = [
{"role": "user", "content": "Create a character card for a grumpy lighthouse keeper."},
]
inputs = tok.apply_chat_template(
messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
).to(model.device)
out = model.generate(
**inputs,
max_new_tokens=8192,
do_sample=True,
temperature=1.0,
top_k=64,
top_p=0.95,
repetition_penalty=1.1,
)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
Cards that include a character_book can be long; if generation cuts off mid-card, raise max_new_tokens. The merged 16-bit weights also serve directly under vLLM (vllm serve aimeri/spoomplesmaxx-cardmaker-v1), again with no system message.
Training Details
Procedure
LoRA fine-tune with Unsloth + TRL SFTTrainer, using the official Granite 4.1 chat template. Loss was computed on the assistant (card) completion only via train_on_responses_only.
LoRA configuration
Table with columns: Setting, Value| Setting | Value |
|---|
Rank r | 16 |
lora_alpha | 22 |
lora_dropout | 0 |
| Target modules | all-linear |
| Rank-stabilized LoRA | enabled |
| Bias | none |
Training hyperparameters
Table with columns: Setting, Value| Setting | Value |
|---|
| Epochs | 2 (848 optimizer steps) |
| Per-device batch size | 1 |
| Gradient accumulation | 8 (effective batch size 8) |
| Max sequence length | 8192 |
| Optimizer | adamw_8bit (β₁ 0.9, β₂ 0.999, ε 1e-8) |
| Learning rate | 1e-4, cosine schedule |
| Warmup steps | 25 |
| Weight decay | 0.001 |
| Max grad norm |
Results
Evaluation loss on the 5% held-out split fell from the base checkpoint to the final model over the two epochs (most of the gain came in the first ~100 steps, with a slow grind afterward):
Table with columns: Checkpoint, Eval loss| Checkpoint | Eval loss |
|---|
Base (step 0, eval_on_start) | 2.234 |
| Step 100 | 1.704 |
| Step 400 | 1.656 |
| Final (step 848) | 1.641 |
Final mean training loss was ~1.57. Total wall-clock training time was ~4.6 hours.
Evaluation
Quality was judged primarily behaviorally rather than by a single metric — eval loss is a weak proxy for card quality on a held-out set this small (~178 rows). A fixed prompt battery probed the behaviors that matter for this task:
- Structure & completeness — clean, parseable cards with all expected fields on easy archetypes.
- Constraint adherence — exact name / age / occupation, and a character's voice actually showing up in
first_mes and mes_example rather than drifting generic.
- Sparse invention — building a full, internally consistent card from a near-empty prompt.
- First-message craft — second-person address to
{{user}}, scene-setting, action formatting, in-voice dialogue, and a natural hand-off.
- Register — antagonist/villain cards produced in-character, with no disclaimers, moralizing, or assistant-voice leakage. This is the main reason the model was trained from a base rather than an instruct checkpoint.
Bias, Risks, and Limitations
- Mature content. This model was trained with a mix of Safe for Work and Not Safe For Work cards, and it may generate objectionable content. Please use discretion when generating new cards.
- Structural validity is not guaranteed. Output is generated text, not schema-validated card JSON. Run it through a parser/validator before importing into SillyTavern.
- Card conventions. Output uses
{{user}} / {{char}} macros and assumes a SillyTavern runtime.
- Single-turn only. This generates a card, not a conversation; it is not itself a roleplay partner.
- Inherited bias. The model carries the biases of both the base model and the curated card sources, including their genre, aesthetic, and demographic skew. "High quality" reflects a subjective curation judgment.
Citation
If you use this model, please reference this repository and the base model.
