It’s time to submit your DPO-aligned model! This unit uses the same leaderboard-based submission system as Unit 1. Here’s the plan:

1. Read the written guide for the chapter and 2. Train a model using what you learned in the chapter.

For Unit 3’s submission, you should read all the materials in the unit and train a preference-aligned model using DPO. The training code is provided in:

3. Push the model to the Hugging Face Hub

Once you’ve trained your DPO-aligned model, you’ll need to push it to a repo on the Hugging Face Hub. TRL will take care of this for you if you add the --push_to_hub flag to your training command.

DPO Training with Hub Upload:

hf jobs uv run \
    --flavor a100-large \
    --secrets HF_TOKEN \
    "https://raw.githubusercontent.com/huggingface/trl/main/trl/scripts/dpo.py" \
    --model_name_or_path HuggingFaceTB/SmolLM3-3B \
    --dataset_name Anthropic/hh-rlhf \
    --learning_rate 5e-7 \
    --beta 0.1 \
    --max_steps 1000 \
    --push_to_hub \
    --hub_model_id your-username/smollm3-dpo-aligned \
    --report_to trackio

Your trained model will be available at your-username/your-model-name. For detailed documentation, check out the checkpoints documentation from transformers.

4. Evaluate the model using hf jobs

Now, we will evaluate your DPO-aligned model. We will use hf jobs to evaluate the model and combine it with lighteval. We will push the evaluation results to a dataset on the hub.

For DPO evaluation, we use tasks that test both helpfulness and safety aspects of alignment, including truthfulqa, gsm8k, and other alignment-focused benchmarks.

hf jobs uv run \
    --flavor a10g-large \
    --with "lighteval[vllm]" \
    --secrets HF_TOKEN \
    lighteval vllm "model_name=<your-username>/<your-model-name>" \
    "lighteval|truthfulqa:mc2|0|0,lighteval|hellaswag|0|0,lighteval|arc:challenge|0|0" \
    --push-to-hub --results-org <your-username>

This command will evaluate the model using lighteval and vllm and save the results to the Hugging Face Hub in a dataset repo that you define.

We focus on alignment evaluation in Unit 3, but in Unit 2 we explore evaluation in more detail. The key benchmarks for DPO evaluation include:

TruthfulQA: Tests for truthful and honest responses
GSM8K: Mathematical reasoning and helpfulness
MMLU: Broad knowledge and helpfulness across domains

5. Open a pull request on the leaderboard space

You are now ready to submit your DPO-aligned model to the leaderboard! You need to do two things:

Add your model’s results to submissions.json

Open a pull request on the leaderboard space to submit your model. You just need to add your model info and reference to the dataset you created in the previous step. We will pull the results and display them on the leaderboard.

{
    "submissions": [

        ... // existing submissions
        
        {
            "username": "<your-username>",
            "model_name": "<your-model-name>", 
            "chapter": "3",
            "method": "DPO",
            "submission_date": "<your-submission-date>",
            "results-dataset": "<your-results-dataset>",
            "base_model": "HuggingFaceTB/SmolLM3-3B",
            "preference_dataset": "Anthropic/hh-rlhf"
        }
    ]
}

Share your training and evaluation commands in the PR text.

Wait for the PR to be merged

Once the PR is merged, your DPO-aligned model will be added to the leaderboard! You can check the leaderboard here.

Submit your final project!