Modelpublic

aimeri/spoomplesmaxx-cardmaker-v2

sourceHugging Faceapache-2.0updated 18d agoView on Hugging Face

0likes334downloads

Model Card

spoomplesmaxx-cardmaker-v2

A full-parameter fine-tune of `aimeri/Ministral-3-14B-Base-2512-text` (the text-only export of `mistralai/Ministral-3-14B-Base-2512`) that turns a short, open-ended prompt into a complete SillyTavern character card. Give it a concept, an archetype, a name and a few constraints, or just a one-liner, and it generates a full V2/V3-style card (description, personality, scenario, first message, example messages, and sometimes a lorebook). This release is training checkpoint ckpt-569.

Model Details

—Developed by: aimeri
—Base model: `aimeri/Ministral-3-14B-Base-2512-text`, the language model of `mistralai/Ministral-3-14B-Base-2512` with the vision tower removed (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.
—Chat template: ChatML (<|im_start|> = id 20, <|im_end|> = id 21 and end-of-sequence), claimed from two unused reserved slots of the Tekken vocabulary; no vocabulary resize.
—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

<TBD>

Training Details

Procedure

Full-parameter supervised fine-tuning (no LoRA) with the Hugging Face Trainer. Each row of `aimeri/st-characters-alpaca-v2` became one ChatML exchange: the instruction (plus input when present) as the user turn and the card as the assistant turn, with no system prompt. Loss was computed on the assistant (card) completion only.

—Weights, gradients and activations in bf16; optimizer states in 8-bit (torchao AdamW8bit) with bf16 stochastic rounding on the weight update, so sub-ULP updates are not lost to round-to-nearest.
—Chunked cross-entropy over the 131k vocabulary (logits never fully materialised) and gradient checkpointing.
—Batch size 1 with no packing or padding; rows longer than the maximum sequence length were dropped, not truncated.
—One GPU (NVIDIA RTX PRO 6000 Blackwell Server Edition (96 GB)) on Google Colab, checkpoints shipped off-box during training.

Training hyperparameters

Setting	Value
Training rows	6773 (0.0M tokens per epoch)
Epochs	2 (569 optimizer steps)
Per-device batch size	1
Gradient accumulation	16 (effective batch size 16)
Max sequence length	16384
Optimizer	torchao AdamW8bit, bf16 stochastic rounding (β₁ 0.9, β₂ 0.999, ε 1e-8)
Learning rate	1e-5, cosine schedule
Warmup	3% of steps
Weight decay	0.01 (matrices only, none on norms)
Max grad norm	1.0
Precision	bf16 (no fp32 master weights)
Seed	1985
Frameworks	Transformers 5.16.1, torchao 0.10.0, PyTorch 2.11.0+cu128, Datasets 4.0.0, Accelerate 1.14.0

Results

Evaluation loss on a 357-row held-out split (5% of the dataset; the in-training eval used a fixed 128-row subset of it):

Checkpoint	Eval loss
Base (step 0, `eval_on_start`)	1.595
Step 200	1.061
Step 400	0.985
Final (step 569)	0.965

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 (357 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.

The battery output for this release is in battery.json.

Bias, Risks, and Limitations

—Mature content. This model was trained on 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, the text-only base and the original Ministral 3 release.