sanjitkverma/UIGEN-X-4B-STD-FFT

STD-FFT (Standard to Diverse Full Fine-Tuning) Model

This repository contains a model trained using STD-FFT methodology on the Tesslate/UIGEN-X-FULL dataset.

Repository Structure

sanjitkverma/UIGEN-X-4B-STD-FFT/
├── final_model/         # Final trained model (also at root for easy access)
├── checkpoints/         # Training checkpoints
│   ├── step-200/       # Checkpoint at step 200
│   ├── step-400/       # Checkpoint at step 400
│   ├── epoch-0.5/      # Checkpoint at epoch 0.5
│   └── epoch-1.0/      # Checkpoint at epoch 1.0
└── inference_samples/   # Generated samples during training
    ├── step_0/         # Initial samples (before training)
    ├── step_200/       # Samples at step 200
    └── step_400/       # Samples at step 400

Training Configuration

• Base Model: unsloth/Qwen3-4B-Thinking-2507
• Dataset: Tesslate/UIGEN-X-FULL
• Training Method: STD-FFT (Standard to Diverse with Full Fine-Tuning)
• Max Sequence Length: 11000
• Batch Size: 1 (with 8x gradient accumulation)
• Checkpoints: Every 200 steps and 0.5 epochs

Usage

To use the final model:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("sanjitkverma/UIGEN-X-4B-STD-FFT")
tokenizer = AutoTokenizer.from_pretrained("sanjitkverma/UIGEN-X-4B-STD-FFT")

To use a specific checkpoint:

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load checkpoint at step 400
model = AutoModelForCausalLM.from_pretrained(
    "sanjitkverma/UIGEN-X-4B-STD-FFT",
    subfolder="checkpoints/step-400"
)
tokenizer = AutoTokenizer.from_pretrained(
    "sanjitkverma/UIGEN-X-4B-STD-FFT",
    subfolder="checkpoints/step-400"
)

Training Details

Training started: 2025-08-18T01:44:18.384004

This model was trained using Modal cloud infrastructure with STD (Standard to Diverse) loss scaling for improved generalization on UI generation tasks.