metadata
language: id
tags:
- mistral
- multi-adapter
- lora
- vision-language
base_model: mistralai/Ministral-3-3B-Base-2512
PAD Model — Multi-Adapter Pack Architecture
Instead of blending parameter matrices together linearly or via TIES-merging, this configuration preserves the distinct, non-destructive weights of three task-oriented LoRA adapter tracks inside a unified model asset structure.
Contained Task Adapters
| Adapter Name | Source Directory Track |
|---|---|
pad1 |
Text Moderation Evaluation (Checkpoint 672) |
pad2 |
Keyword Extraction Logic (Checkpoint 3382) |
pad3 |
Image Visual Age Rating Engine (VLM CPT) |
Dynamic Runtime Usage Pattern
from transformers import AutoModelForImageTextToText, AutoProcessor
from peft import PeftModel
# 1. Load structural framework layers
processor = AutoProcessor.from_pretrained("nuresens/pad_model_adapter_v1")
base_model = AutoModelForImageTextToText.from_pretrained("mistralai/Ministral-3-3B-Base-2512", torch_dtype=torch.float16, device_map="auto")
# 2. Register multi-adapter configuration paths
model = PeftModel.from_pretrained(base_model, "nuresens/pad_model_adapter_v1", adapter_name="pad1")
model.load_adapter("nuresens/pad_model_adapter_v1", adapter_name="pad2", subfolder="pad2")
model.load_adapter("nuresens/pad_model_adapter_v1", adapter_name="pad3", subfolder="pad3")
# 3. Target task context routing dynamically before generating text
model.set_adapter("pad3") # Routes current context to the vision processing matrices explicitly