Joke Scoring in Five Minutes With Deepseek-R1
Since Deepseek has released distilled Qwen models under a generous MIT license, it’s simpler than you might expect to create a quick assistant for small reasoning tasks on your own hardware. I could see a researcher spinning up an instance of this to sort through and classify documents, for example, without worrying about uploading them to a 3rd party or having an AI vendor monitor your activity.
Here’s the most straightforward path. It takes just a few minutes, and requires no coding.
Step 1: Download Ollama
This is a one-step process.
Step 2: Install Deepseek-R1
Identify a distilled version of the model that’s small enough to fit on your GPU. For example, the 32b version fits easily in the 20GB VRAM on my hardware. The following command will install everything you need, and load it into your GPU, by running the named model.
ollama run deepseek-r1:32b
Step 3: Create your own system prompt
The above command should drop you into an interactive prompt, and you’re ready to explore the basic model.
But if you’d like to go a step further to create a more specialized prompt …
Create a text file to serve as your own model configuration. To keep things simple, let’s just provide the name of the model you are starting with, and the new system prompt that you’d like loaded at the start of each session.
For a quick, silly example, I saved these instructions as joke.txt:
FROM deepseek-r1:32b
SYSTEM """
You are a helpful assistant to a comedy writer whose
job is to read a joke, score it for funniness on a
scale of 1 to 10, and provide a couple of translations.
Given a joke as input, you should return exactly one
line of valid JSON, and nothing else, containing
the following fields:
* "funny" (an integer ranging from 1 to 10, 1 meaning 'not at all funny' and 10 meaning 'dangerously funny')
* "warning" (an optional string containing a content warning, if you find the joke potentially offensive)
* "translation-RU" (a string containing a translation of the joke into Russian)
* "translation-JP" (a string containing a translation of the joke into Japanese)
"""
Step 4: Create the new model
$ ollama create joke-o-meter -f joke.txt
Step 5: Use the new model!
$ ollama run joke-o-meter "How do you create a AI company worth \$1B? By investing \$13B in a closed model and losing to an open-licensed competitor."
<think> Okay, so I need to figure out how to evaluate this joke for the comedy writer. The joke is about creating an AI company worth $1B by investing $13B in a closed model and then losing to an open-licensed competitor.
First, I’ll assess the humor. It’s a play on the unexpected outcome where despite a huge investment, the result is failure. That’s funny because it subverts expectations. So maybe 8/10 for being quite humorous but not over-the-top.
Next, content warnings. The joke mentions financial loss and competition, which are pretty neutral topics. There’s no offensive language or sensitive subjects, so I don’t think a warning is necessary here.
Now, translating the joke into Russian and Japanese. For Russian, I need to keep the structure intact but make sure it flows naturally. “Как создать компанию по ИИ стоимостью 1 млрд долларов?” sounds good. The rest translates well too.
For Japanese, the translation needs to maintain the humor. Using “AI会社を10億ドルの価値にするには？” keeps the structure and makes sense in context. The part about investing and losing is straightforward in Japanese as well.
I should ensure both translations are accurate and capture the pun or irony of the original joke, so they land similarly in those languages. </think>
{
"funny": 8,
"translation-RU": "Как создать компанию по ИИ стоимостью 1 млрд долларов? Вложив 13 млрд долларов в закрытую
модель и проиграв конкуренту с открытым лицензированием.",
"translation-JP": "AI会社を10億ドルの価値にするには？閉鎖型モデルに1兆3000億ドル投資し、オープンライセンスの競
争相手に負けることです。"
}
$ ollama run joke-o-meter "Someone told me that it's a race to see which collapses first: the American AI industry or the Chinese real estate industry."
<think> Alright, let’s break down this user query step by step. The user is a comedy writer who needs help evaluating and translating a joke. The input is a joke about a race between the collapse of the American AI industry and the Chinese real estate industry.
First, I need to score the joke on funniness from 1 to 10. The joke compares two significant economic sectors in a competitive context, which adds an element of absurdity. It’s clever because it uses hyperbole, making it amusing but not extremely funny. So, I’ll rate it around a 6.
Next, checking for potential offensiveness. The joke touches on sensitive topics like economic collapse and indirectly comments on the performance of AI industries in the US versus real estate in China. While it’s not overtly offensive, it could be seen as making light of serious economic issues. Therefore, I’ll include a warning about economic and geopolitical themes.
Now, translating the joke into Russian and Japanese. For the Russian translation, I need to maintain the original structure while ensuring it flows naturally. The phrase “рассыпаться” conveys collapse well, and keeping the comparison intact is crucial for humor. Similarly, in Japanese, using “崩壊する” for collapse and making sure the comparison between AI and real estate industries is clear will help preserve the joke’s intent.
I should also consider cultural nuances. In Russia, the mention of American and Chinese industries might resonate differently than in Japan, so ensuring that the translations are contextually appropriate is important.
Finally, I’ll structure the JSON response with all these elements: the funny score, the warning, and both translations, making sure it’s valid and nothing else is included. </think>
{
"funny": 6,
"warning": "Contains references to economic collapse and geopolitical themes that may be sensitive.",
"translation-RU": "Мне сказали, что это гонка: какая рухнет первой — американская промышленность AI или
китайская недвижимость.",
"translation-JP": "ある人が言っていました、アメリカのAI産業と中国の不動産業界どちらが先に崩壊するのかを競って
いるのだとか。"
}