---
library_name: gguf
tags:
- phi-4
- danbooru
- art-tagger
- quantized
- text-generation
base_model:
- USS-Inferprise/Phi4-Mini-Prose2Tags-4B
license: mit
pipeline_tag: text-generation
---
Quants of (https://huggingface.co/USS-Inferprise/Phi4-Mini-Prose2Tags-4B)

quantized_by: USS-Inferprise

# We also include a concept for a ComfyUI custom node for applying this model in a workflow.

Original Model Card Follows:

# Phi4-Mini-Prose2Tags-4B

This model is a specialized fine-tune designed to translate natural language prose descriptions into structured **Danbooru-style tags**. It is intended to bridge the gap between human-readable image captions and the tag-based prompting systems used by many latent diffusion models.

## Model Details

- **Developed by:** USS-Inferprise
- **Model Name:** Phi4-Mini-Prose2Tags-4B
- **Base Model:** [huihui-ai/Phi-4-mini-instruct-abliterated](https://huggingface.co/huihui-ai/Phi-4-mini-instruct-abliterated)
- **Training Architecture:** LoRA (Low-Rank Adaptation)
- **Merging Method:** Linear Merge (via Mergekit)
- **Primary Task:** Prose-to-Tag Translation

## Training Methodology

### Dataset Construction
The training data ([USS-Inferprise/Phi4-Mini-Prose2Tags-4B-Raw-Training-Data](https://huggingface.co/datasets/USS-Inferprise/Phi4-Mini-Prose2Tags-4B-Raw-Training-Data)) was generated using a synthetic pipeline:
1. **Source Images:** 100,000 images sourced from `laion/conceptual-captions-12m-webdataset`.
2. **Prose Generation:** Images were described using **QwenVL**.
3. **Tag Generation:** Images were tagged using **WD 1.3**.
4. **Pairing:** The resulting QwenVL descriptions and WD 1.3 tags were paired to create the final training instruction set.

## ⚠️ Safety & Content Note

> [!IMPORTANT]
> This model was trained exclusively on a curated subset of data intended for general audiences. **No explicit, NSFW, or adult-oriented tags** were included in the training dataset (`Prose2Tags-4B-Raw-Training-Data`). 
> 
> While the base model (`Phi-4-mini-instruct-abliterated`) has been modified to reduce certain refusals, this specific fine-tune is designed for clean, descriptive tagging. It may not recognize or accurately generate tags related to explicit content. If it can... it didn't learn it from us.

### Training Process
- **Library:** [Unsloth](https://github.com/unslothai/unsloth)
- **Hardware:** NVIDIA L40S
- **Epochs:** 1
- **Method:** LoRA fine-tuning merged into the base model using a Linear merge strategy.

## Evaluation & Testing
Testing was performed on 100 images excluded from the training set. To ensure the model generalizes well across different captioning styles, the test inputs used **gokaygokay/Florence-2-SD3-Captioner** instead of the training-source QwenVL.

Detailed test outputs can be found here: [USS-Inferprise/Phi4-Mini-P2T-4B-Testing](https://huggingface.co/datasets/USS-Inferprise/Phi4-Mini-P2T-4B-Testing).

## Proper Prompt Format

**Warning:** You must strictly follow the prompt format below. Failure to do so may result in the model reverting to the standard Phi-4-Mini helpful persona rather than generating tags.

```text
<|user|>
You are a Danbooru tag translator.
{prose_input}<|end|>
<|assistant|>