Abuja-01

This model is a fine-tuned version of Qwen/Qwen2-0.5B-Instruct. It has been trained using TRL with ORPO (Odds Ratio Preference Optimization).

Model Details

• Base Model: Qwen/Qwen2-0.5B-Instruct
• Training Method: ORPO (Odds Ratio Preference Optimization)
• Training Dataset: trl-lib/ultrafeedback_binarized
• Training Time: 1 hour 32 minutes
• Hardware: Single GPU

Training Metrics

• Training Loss: 6.386
• Train Samples per Second: 11.152
• Train Steps per Second: 0.697
• Final Epoch: 1.0

Quick Start (running on CPU)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
import logging

logging.basicConfig(level=logging.INFO)

def load_model():
    """Load the model and tokenizer."""
    logger.info("Loading model...")
    model = AutoModelForCausalLM.from_pretrained("iben/Abuja-01", trust_remote_code=True)
    tokenizer = AutoTokenizer.from_pretrained("iben/Abuja-01", trust_remote_code=True)
    return model, tokenizer

View the full code on GitHub Gist

<script src="https://gist.github.com/ibeninc/8ddc165576dee140cb118265b34335c5.js"></script>

Training Details

This model was trained using ORPO (Odds Ratio Preference Optimization), a method that doesn't require a reference model. The training configuration included:

• Learning Rate: 1e-5
• Batch Size: 4
• Gradient Accumulation Steps: 4
• Training Epochs: 1

Framework Versions

• TRL: 0.13.0
• Transformers: 4.48.1
• PyTorch: 2.5.1+cu121
• Datasets: 3.2.0
• Tokenizers: 0.21.0