LLM2025_Advanced_DPO_5

This repository provides a DPO-fine-tuned model based on rokugatsu/LLM2025_Advanced_5 using trl.DPOTrainer.

This model has undergone Direct Preference Optimization (DPO) to align with human preferences, using trajectories from agent-based tasks.

Training Objective

This model was fine-tuned using DPO to improve multi-turn agent task performance by learning preferences from the u-10bei/sft_alfworld_trajectory_dataset_v2 dataset. The DPO training process aims to increase the likelihood of generating 'chosen' responses and decrease the likelihood of 'rejected' responses for given prompts.

Training Configuration (DPO)

• Base SFT Model: rokugatsu/LLM2025_Advanced_5
• DPO Dataset: u-10bei/sft_alfworld_trajectory_dataset_v2
• DPO Method: Direct Preference Optimization (DPO)
• Max sequence length: 2048
• Epochs: 0.25
• Learning rate: 2e-06
• Beta parameter (DPO loss): 0.1

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

model_id = "rokugatsu/LLM2025_Advanced_DPO_5"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16, # Use bfloat16 if your GPU supports it
    device_map="auto",
)
# The model is already merged, so no need for PeftModel.from_pretrained(model, adapter)

# Example for inference (assuming you have a chat_template)
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "What is the capital of France?"}
]
input_ids = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)

outputs = model.generate(input_ids, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Sources & Terms (IMPORTANT)

Training data: u-10bei/sft_alfworld_trajectory_dataset_v2

Dataset License: MIT License. This dataset is used and distributed under the terms of the MIT License. Compliance: Users must comply with the MIT license (including copyright notice) and the base model's original terms of use.