You can deploy this model and start earning money today!

ErMayureshKumar/
ai-fashion-stylist-355m

Name: ai-fashion-stylist-355m
Availability: InStock

Views

No views yet

First Generation AI Fashion Stylist 355M

A from-scratch 355M parameter GPT-style language model specialized as a local-first AI fashion stylist.

The model was developed as a research project exploring:

Small language models
Instruction tuning
Domain-specific adaptation
Synthetic data distillation
Catastrophic forgetting mitigation
Local/offline deployment

Unlike cloud-dependent assistants, this model is designed to run efficiently on consumer hardware while providing useful fashion guidance and maintaining general conversational ability.

Model Overview

The training pipeline consists of three stages:

Pretraining
- From-scratch GPT-style language model (~355M parameters)
- Trained on approximately 1.18B tokens of general text
General Instruction Tuning
- Trained on instruction-following and chat datasets
- Improves conversational ability and instruction adherence
Fashion Domain specific Tuning
- Trained on curated fashion datasets and synthetic fashion Q&A
- Uses rehearsal-based training to preserve general abilities
- Focuses on outfit recommendations, wardrobe planning, and fashion advice etc.,

The final model acts as a lightweight AI stylist while retaining much of its general assistant behavior.

Intended Uses

Primary Uses

Fashion advice chatbot
Outfit recommendations
Wardrobe planning
Capsule wardrobe creation
Office wear suggestions
Wedding and festival outfit guidance
Indian fashion guidance
Educational and research experiments
Local-first AI assistant research

Out-of-Scope Uses

This model is not intended for:

Medical advice
Mental health advice
Legal advice
Financial advice
Body image assessment
Professional fashion consulting
High-risk decision making
Commercial production systems

Training Data

Pretraining

General-purpose text corpus - cosmopedia used to train the base language model.

Approximate scale:

~1.18B tokens

The objective was to create a domain-agnostic language model before specialization.

Stage 1 Instruction Tuning

Datasets include:

Alpaca
Dolly
OpenAssistant
UltraChat

Approximate scale:

~390M tokens

Purpose:

Improve instruction following
Improve helpfulness
Improve conversational behavior

Stage 2 Fashion Tuning

Fashion-specific training data includes:

nreimers/fashion-dataset
Teacher-generated synthetic fashion instruction datasets

Topics include:

Weddings
Office wear
Festivals
Casual outfits
Travel clothing
Capsule wardrobes
Indian ethnic wear
Body-type-aware styling
Budget-conscious fashion etc.

Approximate scale:

~40M tokens

Methodology

The fashion adaptation stage uses a forgetting-aware training strategy.

Key techniques include:

Synthetic data distillation
Rehearsal training
Mixed-domain batches
General instruction retention
Domain-specific supervised fine-tuning

The objective is to improve fashion expertise while minimizing degradation of general assistant capabilities.

Hardware

All stages of training were completed on a single consumer laptop.

Training Hardware

Machine: Alienware 16 Aurora AC16250
GPU: NVIDIA RTX 5060 Laptop GPU (8GB VRAM)
CPU: Intel Core i7-240H
RAM: 16GB
Storage: NVMe SSD

Total training volume:

Pretraining: ~1.18B tokens
Stage 1: ~0.39B tokens
Stage 2: ~0.04B tokens

Total:

~1.6B tokens processed

This project demonstrates that a complete pretraining + domain specialization pipeline is feasible on modest consumer hardware.

Evaluation

Evaluation focused on two dimensions:

General Capability

Measured using:

Held-out instruction tasks
Conversational prompts
General QA

Goal:

Preserve Stage-1 assistant behavior

Fashion Capability

Measured using:

Synthetic reference comparisons
Human/model-judge comparisons

Observed outcome:

Significant improvement in fashion-related responses
Minimal degradation in general conversational ability

Limitations

Users should be aware of several limitations.

Text-only

The model cannot process:

Outfit photos
Body images
Wardrobe images
Product images

Small Model Capacity

Compared with larger models:

Less reasoning depth
Less personalization
Limited long-context performance

Biases

The model may inherit biases from:

Training datasets
Synthetic data generation
Cultural assumptions

Trend Awareness

The model does not have access to:

Live fashion trends
Real-time brand information
Current product catalogs

Professional Advice

The model is not a professional stylist and should not replace expert consultation.

Usage

Installation

pip install transformers torch

Loading the Model

python

from transformers import AutoTokenizer
from transformers import AutoModelForCausalLM
import os
import torch

MODEL_ID = "ErMayureshKumar/ai-fashion-stylist-355m"

tokenizer = AutoTokenizer.from_pretrained("gpt2")

special_tokens = {
    "additional_special_tokens": [
        "<|user|>",
        "<|assistant|>"
    ]
}

tokenizer.add_special_tokens(special_tokens)

model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float16,
    device_map="auto"
)

prompt = """
Suggest an outfit for a summer wedding in Mumbai
for a 26-year-old woman.
"""

def chat(prompt):

    text = f"<|user|>\n{prompt}\n\n<|assistant|>\n"
    inputs = tokenizer(
        text,
        return_tensors="pt"
    ).to(model.device)

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=512,
            temperature=0.7,
            top_p=0.9,
            top_k=50,
            do_sample=True,
            repetition_penalty=1.1,
            pad_token_id=tokenizer.eos_token_id,
            eos_token_id=tokenizer.eos_token_id,
        )
    result = tokenizer.decode(
        outputs[0],
        skip_special_tokens=False
    )
    return result

print(chat(prompt))

Example Prompts

Wedding

I am a 30-year-old man attending a summer wedding. Suggest a complete outfit.

Office

Explain business casual attire.

Capsule Wardrobe

What is a capsule wardrobe?

General Chat

What should I consider before changing careers?

Future Work

Planned future directions include:

Multimodal fashion understanding
Outfit image analysis
Fashion retrieval systems
Personalized style profiles
Fashion knowledge graphs
Additional languages
Enhanced reasoning
Mobile deployment optimization

License

Educational and Research Use License

Permission is granted to use, copy, modify, and distribute this model for educational, academic, and non-commercial research purposes only.

Commercial use, commercial redistribution, deployment in paid services, integration into commercial products, or any revenue-generating use is prohibited without prior written permission from the copyright holder.

THE MODEL IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.

Citation

bibtex

@misc{aifashionstylist355m,
  title={AI Fashion Stylist 355M},
  author={Er Mayureshkumar},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/ErMayureshKumar/ai-fashion-stylist-355m}
}

Acknowledgements

This project builds upon the open-source ecosystem including:

Hugging Face Transformers
PyTorch
cosmopedia
Alpaca
Dolly
OpenAssistant
UltraChat
Fashion-domain datasets (nreimers/fashion-dataset)
synthetic distillation methodologies

Intended Use Statement

This model and repository are released solely for educational and non-commercial research purposes. The model should be viewed as a research artifact demonstrating local-first domain adaptation of a small language model into a fashion-focused assistant.