voidx-decider-12b
a specialized LoRA adapter for classification and decision-making
2 min read
Built on google/gemma-4-12B-it by upfall-co, voidx-decider-12b is trained to output calibrated probabilities over answer choices instead of generating free-form text. It evaluates the probability of each label token at a fixed answer position and supports both English and Korean text. The adapter is tailored for classification, ontology mapping, and state evaluation in data processing pipelines. Running it requires the base Gemma 4 12B model and the PEFT library (a lightweight wrapper on top of a base model for efficient fine-tuning).
Key features
- Calibrated probabilities: reads label token probabilities at the answer position rather than generating text.
- Multilingual support: full support for prompts in both English and Korean.
- Efficient architecture: a LoRA adapter (r=32, alpha=64) built on top of a 12B language model.
- Multiple formats: available as a merged Q4_K_M GGUF build alongside the original weights.
What you can do
- Pipeline integration: plug it into automated document and text categorization services.
- Ontology mapping: bind entities to taxonomy elements and classification schemas.
- State evaluation: handle state tracking and logical decision-making in bots and applications.
- Local deployment: spin up the GGUF build via
llama-serverfor fast structured-query processing.
Example
Loading the model with the PEFT library in Python:
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained(
"google/gemma-4-12B-it", revision="707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7",
torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(base, "upfall-co/voidx-decider-12b")
tok = AutoTokenizer.from_pretrained("google/gemma-4-12B-it",
revision="707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7")
Read more
DetailsValidation metrics: 0.822 accuracy, 0.028 ECE.
Training: 1 epoch on ~14.5k examples using a single H100 80GB GPU in 90 minutes.
Limitations: vulnerable to direct prompt injections (selects the injected option 98% of the time), tends to pick a category in IPTC ontologies even when none fit, and is sensitive to option ordering.
License: Apache-2.0.










Comments