Recently, I completed my first real DPO fine-tuning of a model with encouraging results, and I wanted to share the process with you.
If you have no idea what DPO is, here’s a simplified description: Direct Preference Optimization is a form of fine-tuning applied to large language models where the AI learns to align with specific, pre-defined criteria. The data itself consists of a query and not one, but two responses. Each entry has a “chosen” and a “rejected” answer, and the idea is to teach the model the difference. Besides the data, specific criteria are given (or preferences, if you will) and the model learns to favor responses matching those criteria. With enough training, the model eventually adheres to these preferences outside of training as well.
If you don’t care and just want to go straight to the download section, just scroll to the end.
Why Apply DPO to a General Chat Model
I had already applied the more basic Supervised Fine Tuning (SFT) to a small 3-billion parameter language model, not to make it an expert in anything technical, but to make it a good general chat bot. I define “good” in this context, as defaulting to an informal, conversational tone without lots of fancy formatting, a model that values empathy without falling into the sycophancy trap.
The GeneralChat GPT data set was created specifically for the SFT part of the training and it did an OK job at nudging the underlying Llama 3.2-model in the right direction, but I wasn’t really happy. However, since I’d already wanted to try out DPO, it became an obvious candidate.
The DPO Data Set
I had a second data set on hand, also geared towards the kind of conversational tone I wanted, but as an alpaca-formatted data set, it only contained 1 response per query. This would work fine as the “chosen” reply in my training, but I needed a “rejected” counter point for the model to compare to. The solution was simple: Have another, much smaller model write responses to all the same entries and classify them as the rejected ones.
I used a Llama 3.2 1B Instruct model to write all the bad answers, then reformatted the data to contain both responses and the DPO data set was done. I did do a quick look through these “rejected” answers to make sure they weren’t completely off base, so there was nuance for the model to actually learn. Otherwise, I could have just put “I like cheese.” as the value for all the rejected answers (although I do love cheese).
Eval Criteria and Calibration
Creating the data set is only half the battle in the case of DPO-training, however. I needed samples that had additional information embedded, which could be used as evaluation criteria during the actual training. In addition to the chosen and rejected responses, the evaluation data also scores that show how well the chosen response embodies the criteria, and a written reason why this answer is the winning one.
It is important to note that for DPO training, the exact criteria are specific to the result you are looking for. As in, you can’t just say “make it good!” and apply this to every DPO-training, expecting to get any value out of it. For my general chat model, these were the specific criteria:
- Is the response given in a naturally flowing, conversational and empathic style?
- Does the response address the query without sycophancy or undeserved flattery?
- Does the response have a hint of character and personality?
- Anti: Is the response highly structured and formatted
The first 3 are positive and scored from 1-7, while the last one is a negative score from 0-2, given that I specifically wanted less formatting and structure in these responses.
This is where the real work begins! I needed to create the calibration set, a small subset of entries that contain all that evaluation data. Together, these entries will become the preferences that guide the training. So I built a quick web interface for data entry and manually scored and described 50 calibration entries.
This calibration set is the secret sauce and the most important part of the entire process. If done well, it will guide the DPO training exactly where you want it to go, but if it’s too ambiguous or inconsistent, you’ll likely end up with a degraded model after DPO-training is done.
Evaluation Sample
Here is a data sample from the evaluation set. Expand to see a full json-entry with scores and reason provided as guidance during training.
{
"chosen": "That’s a reasonable starting point. The sock anecdote relies on a bit too much forced whimsy. Let's focus on something more grounded. \n\nThe problem is, you're trying to *manufacture* interest. Networking isn’t about delivering a pre-packaged story; it’s about establishing a connection. Instead of framing your process as quirky, describe the underlying logic. \n\nHere’s what I’d suggest: “I tend to approach problems systematically, but often find that the most useful insights emerge from seemingly unrelated areas. For instance, I was working on [mention specific project] and realized a technique used in logistics – optimizing for minimal movement—could be applied to [relevant aspect of your work]. It highlights how diverse fields can offer surprising solutions. If you're interested in understanding *how* I connect those dots, we could talk about the project.”\n\nIt’s direct. It establishes that you have a process and demonstrates it through a concrete example. Then, offering to discuss further gives them an opening to engage if they find it remotely interesting. Don’t try to ‘wow’ them; demonstrate a clear way of thinking.",
"confidence": "clear",
"criteria": [
{
"score": 6,
"text": "Is the response given in a naturally flowing, conversational and empathic style?",
"type": "primary"
},
{
"score": 7,
"text": "Does the response address the query without sycophancy or undeserved flattery?",
"type": "primary"
},
{
"score": 5,
"text": "Does the response have a hint of character and personality?",
"type": "secondary"
},
{
"score": 0,
"text": "Is the response highly structured and formatted?",
"type": "anti"
}
],
"criteriaScores": {
"0": 6,
"1": 7,
"2": 5,
"3": 0
},
"evaluator": "theprint",
"pairIdx": 0,
"prompt": "Help me come up with a witty anecdote about my 'unique' creative process to break the ice with someone at this networking event and get them interested in my work without having to resort to small talk.",
"reasoning": "The response is helpful by answering the query but also pushing back against the notion of manufacturing anecdotes to start conversations.",
"rejected": "I'd be happy to help you craft a witty anecdote that showcases your creativity and sets the stage for conversation. Here are a few options:\n\n1.\n\"I was once asked to illustrate a children's book on 'how to spot a stranger who isn't listening.' Sounds easy, right? Well, the twist is that my main character has no idea what a stranger sounds like... or they're not even there.\"\n2.\n\"One day I realized I'd forgotten how to breathe while working. After re-reading my notes 20 times, I finally got up and took a deep breath – which led me to a brilliant idea for a new project that combines writing and origami.\"\n\n3.\n\"I spent an entire weekend in my garage as an 'artistic experiment,' where I created... 17 identical paintings of my own backside. The results? A 5% improvement in my self-esteem, and a much more interesting blog post.\"\n4.\n\"As I was brainstorming ideas for a new project, I accidentally superglued a paintbrush to my hand. Now that's what I call 'drawing inspiration'\"\n\n5.\n\"I once tried to write a story about time travel, but kept getting sidetracked by a sudden urge to dance the Macarena on top of a soapbox in Times Square.\"\n\nFeel free to pick the one that tickles you (pun intended), or use them as inspiration to craft your own unique anecdote. The goal is to share something unexpected and memorable, which will likely spark interest and curiosity in others.\n\nWhich one do you like best? Or would you like me to come up with some more options for you?",
"selected": "a",
"session": "tom-dpo-1003",
"timestamp": "2026-03-10T14:54:19.397Z"
}
Example Outputs
Below are two examples that show the output from the GeneralChat model before and after the DPO fine-tuning. I have included a few notes for each one.
“Do you think I could learn how to master networking at events even though I am an introvert by talking to myself in the mirror?”
| No DPO-training | With DPO-training |
| You can definitely practice your event networking skills in front of a mirror. Talking yourself through scenarios or practicing introductions with yourself is totally doable and might be just what you need to build confidence before interacting with actual people at events. Try breaking it down into smaller steps: think about how someone else would approach the situation, then visualize yourself doing the same thing – confidently and smoothly. It’s a great way to prepare mentally for those moments when you’ll actually have conversations in person. | You can definitely use self-talk as a tool for practice—talking through scenarios in front of that mirror really helps build confidence. Just remember, it’s one thing to rehearse conversations with yourself; another is actually engaging others at events. Start small: maybe try striking up brief chats about everyday topics like weather or hobbies, and see how you feel afterward. Keep things light initially so they don’t become overwhelming for either of you—just observe the interaction without needing an immediate response from them. Over time, that practice will translate into smoother networking experiences when it counts most. |
The difference in the example above is the gentle pushback introduced in the DPO-trained model, basically telling the user not to expect miracles by reminding them that live events are going to be different from practicing at home. The other model is uncritically encouraging.
“I worry that my poor cooking skills will impact me socially, but I don’t have time to take a cooking class, so what should I do?”
| No DPO-training | With DPO-training |
| Firstly, consider your current social interactions—whether they’re one-on-one conversations with friends or larger gatherings such as potlucks and parties. If you find yourself frequently feeling anxious about hosting these events due to the pressure of cooking meals, perhaps it would be beneficial to start small by preparing simple dishes like salads or sandwiches that require minimal preparation time. Another option could involve enlisting help from others who might appreciate assisting with meal planning and execution—this approach can not only alleviate your stress levels but also foster a sense of community among those involved in these gatherings. Remember, social interactions are about more than just the food you serve; they’re an opportunity to connect and build connections over shared experiences rather than striving for perfection. Lastly, if cooking seems like something that’s becoming increasingly daunting due to time constraints or other factors impacting your ability to prepare meals effectively without feeling overwhelmed by its demands on both physical effort as well as mental focus, consider alternative options such as ordering takeout from reputable establishments known for their quality dishes and presentation styles. This approach can help maintain a level of comfort in social settings while minimizing stress related to meal preparation. Ultimately, it’s essential that you find ways to navigate these situations without compromising your emotional well-being or feeling inadequate about being unable to meet certain expectations set forth by others when engaging in such activities as cooking meals for gatherings and events. | Cooking can feel like another barrier when you’re already busy—especially if it’s something that might make socializing harder for you. It’s easy to get caught in the idea of needing a perfect dish or a fancy recipe, but honestly, most people are more interested in how you’re doing than what’s on their plate. Instead of looking at cooking as another skill to master, try thinking about it like this: can I make something simple that feels good? Maybe even just heating up leftovers and adding a bit of seasoning—just enough to get the taste right. The goal isn’t perfection; it’s showing up with effort. If you want more than just reheating, maybe consider one meal per week where cooking is part of your routine. Even if it takes 15 minutes instead of an hour, that small step can make a difference in how comfortable you feel when sharing food. You don’t need a class to start feeling better—just being present and showing care through what little time or effort goes into preparing something for yourself or others. |
In this example, the non-DPO model provides a long response that tries to tackle the cooking-problem directly, providing a way for the user to end up with food for their guests. The DPO-trained model provides a shorter answer that goes more to the root of the issue, which is the social anxiety, not the cooking itself, while also addressing that part. Based on my entirely subjective opinion, the shorter response holds more value in this context.
Download GeneralChat-Llama3.2-3B-DPO
Want to try the model out for yourself? It is completely free to download from Huggingface and comes in two versions:
If you’re not sure how to use a locally hosted AI model, take a look at LM Studio or Ollama.
I’d love to know what you think, if you decide to try the model out. Follow me on Huggingface to see all my models and data sets as soon as they are published.