This Kaggle guide was originally made for JanitorAI (As shown from Step 5 and onwards). But this can also be used for any other frontend you might like to use.
SNOWDROP
Things you'll need for this guide:
A phone number (VOIP numbers are possible, so you can look into free phone number websites!)
Step 1. Go to nGrok's website at https://ngrok.com/. We're going to create an authentication token. This will be important for later.
Create an account 
1.1 You can skip setting up Multi-factor Authentication.
1.2 Once on the main menu. Go to "Your Authtoken". On the top-left of your screen.
1.3 Next. Just copy your Authtoken and save it somewhere for now. We will use this later.
Step 2: Setting up Kaggle.
Important: For this step you need any phone number!
2.1 Head on over to https://www.kaggle.com/# and create your account.
2.2 Once you have made your account, go to your settings at https://www.kaggle.com/settings, or by clicking your icon at the top right of your screen.
Here we'll be using your Phone Number to gain access to more features!
Step 3. Using the Kaggle!
Important: For this step you will need your Auth key from nGrok!
3.1 Go to this kaggle link: https://www.kaggle.com/code/divinesinner/koboldcpp-guide-in-comment/notebook
And then you want to press Copy & Edit
3.2 If everything went right, you should be on this screen!
3.3 First off, you want to hover over "Add-ons" in the top left, and click on Secrets!
3.4 This will open up a menu on the side that will allow you to "Add secret"
Click on it.
3.5 It will open up a New Secret menu. Put in the Label section: "ngrok-auth". And put the Authtoken you got at https://ngrok.com/ in the Value section. See Example. Then save!
3.6 The menu should now look like this. Make sure you have the Secret enabled with the checkmark!
3.7 Next, we're going back to the top-left! But only this time, we're gonna change some settings!
3.8 Next, scroll all the way down until you see the final node!
3.9 You're going to want to edit the 32000 into 16000, this will make the model run better!
YOU'RE NOW DONE SETTING UP THE KAGGLE! 🎉
Step 4. Starting the Kaggle!
4.1 Click on Run All!
4.2 Important. Start playing the audio at the top of then notebook once it's visible! This is to keep everything running!
4.3 Important. Make sure that the model got downloaded properly! Don't worry about the error, but make sure it's at 100%!
4.4 Copy your API Link by scrolling down a bit! If it's not there, make sure you set your Secret at step 3.6 and that the model is 100% downloaded at step 4.3!
NOTE
THIS API URL CHANGES EVERYTIME YOU RESTART THE KAGGLE, THIS MEANS YOU NEED TO COPY IT AGAIN EVERYTIME YOU USE THE KAGGLE!
4.5 Scroll all the way down. And wait for KoboldCPP to start properly! Wait until you see the message in the screenshot appear!
You have now succesfully launched the model!
NOTE: DO NOT CLOSE OR RESTART THE KAGGLE, KEEP IT OPEN TO KEEP IT RUNNING!
Step 5. Onto https://janitorai.com.
Go into a chat and into your Proxy menu!
5.1 Paste the API Link you got earlier into the API URL box and then click on "Add /chat/completions to url" like in the screenshot!
You do NOT need a model name
You do NOT need an API KEY (But this box cannot be empty, so put anything you want inside of it!)
5.2 REFRESH!!!
5.2 Now, you're finished! Try sending a message!
Snowdrop FAQ
Since I was asked. Here are some recommended settings:
Temperature = 1
Token Limit = 0
Context Size = 16000
Here are also some prompts that you can try out on Snowdrop!
https://huggingface.co/trashpanda-org/prompts-and-presets/blob/main/sev's%20prompt.txt
https://rentry.org/notesfromunderground
Want to try hosting your own model, without relying on Kaggle? Try this guide by Blitzen! https://waiki.trashpanda.land/guides:self_hosting_local_kobold
Short FAQ:
What is Snowdrop
Snowdrop is a model finetuned for roleplay specifically. This means this unlike models like Deepseek and Gemini which is all-purpose, this model is made purely for roleplay!
Is it better than Deepseek?
Because Snowdrop is a smaller finetuned model in comparison to a massive generalist model, it can depend of preference. Snowdrop was MADE for roleplay, this means that it'll likely outperform Deepseek in aspects! But Deepseek has a massive database, this means it has a lot more data to work with and might produce better results!
Is it free?
Yes! Through Kaggle you can use up to 30 hours per week of Snowdrop. This is per account, but as I mentioned in the guide. You CAN use VOIP phone numbers. (This means that if you find the right website... you can just yoink a phone number from there for a new account... You didn't hear that from me though)
More Models!
First of all! This is a continuation of my Snowdrop guide at! You will atleast need to have followed Step 1, 2 and 3 of that guide to set up your Kaggle account!
There is also a more concise Rentry version of this guide if you already know how to work around the Kaggle at
https://rentry.org/severian#selecting-a-model-quant-to-run-on-kobold-kagglegoogle-colab!
So to get things in order. Here is a checklist!
Checklist:
- Obtained your Auth Key at nGrok Step 1
- Verified your phone number on Kaggle (tep 2
- Put your Auth Key in your Secrets menu. Step 3
- Selected the correct accelerator. Step 3.7
- OPTIONAL. THIS MIGHT DEPEND ON THE MODEL, BUT IS USUALLY IDEAL.
Changed your context size to 16000 Step 3.8
ALRIGHT. LET'S BEGIN!
Step 1. Learning more about Kaggle!
So remember how we set our accelerator at GPU T4 x2? This means when we're running our notebook (code), we're using 2 Nvidia T4 GPUs. Each GPU has access to 16 GB VRAM (Video RAM) , so that means we have 32 GB VRAM in total!
Of course. This doesn't mean we should be using all those 32 GBs of VRAM. So we'll be reaching for models that use LESS than 30 GBs of VRAM.
Step 2. Finding a model!
THIS IS A WORK IN PROGRESS!
We're still working on a list of recommended models! For now, take these!
https://colab.research.google.com/drive/1l_wRGeD-LnRl3VtZHDc7epW_XW0nJvew
HibikiAss' colab contains a bunch of models you could use + short reviews for them. These are models recommended for colab which has less VRAM than Kaggle, so there might be better out there 🥺 (again. Work in progress friends.)
https://huggingface.co/bartowski
https://huggingface.co/mradermacher
These are two quanters that you should look into, these will have a bunch of models that you can use on the Kaggle!
https://www.notion.so/playwithpepper/1f392d900248803f86c2c51c73f92a0b?v=1f392d90024880d59e33000cc1b15175
Pepper's rentry based on Featherless models with short reviews!
https://rentry.org/anathem
Myscell's rentry, MUST CHECK OUT! ❤️
Step 2.1. Alright! Let's say I've found a model I want to use! For this example I'll be using Sao10K/Llama-3.3-70B-Vulpecula-r1.
First off, we're going to calculate if we can use this model. For this, we'll be using https://huggingface.co/spaces/NyxKrage/LLM-Model-VRAM-Calculator.
Step 2.2 Let's start off by filling the information that we've collected! First off, our GPU we don't need to fill in.
We have 32 GB of VRAM!
Step 2.3 Next off. We'll put in our model. You can find the model name on the Hugging Face page!
Let's fill it in.
Step 2.4 Filling in our context size! For this example we're using a context size of 16384
NOW. Let's submit this!
Aaaand...
Our Total Size (GB) is 32 (And preferably even lower than that). That means that this quant DOES NOT FIT.
Step 2.6 Okay. Okay. So let's try that again! Let's change the quant and see what happens! Let's try IQ2_M! Which...
Is BELOW a total size of 30 GB. Which means it FITS.
Great! Now that we've found a quant that fits in Kaggle. Let's go look for that!
Step 3. Acquiring the model! Let's take a look at the HF page of our chosen model and see if it has Quants. Look for the Quantizations tab under the Finetunes for our selected model!
Step 3.1. You should appear on a page that looks like this. Let's look for the Quanters that I spoke about earlier.
Step 3.2, Locate the quant we know will work, and copy the link!
Step 4. TIME TO KAGGLE. Yeah that's right. Time to actually use Kaggle.
Go to https://www.kaggle.com/code/divinesinner/koboldcpp-guide-in-comment, or preferably your copy of this Kaggle that has everything set-up.
(See Checklist at step 1)
Step 4.1. Scroll down until you see This
Step 4.2 We're going to replace the URL here, with our own. IMPORTANT: ONLY REPLACE UP TO THE .gguf
Like this.
Step 5. TIME TO RUN ALL. Everything should be the same. You will know if you pasted the right model name if it's correctly downloading it here!
Step 5.1 Follow the rest of the steps of my other guide from Step 4.2!
SO IF EVERYTHING WENT CORRECTLY.
THIS GUIDE WAS MADE BY NAEN. PLEASE DO NOT COPY OR STEAL MY WORK. IF NEED TO CONTACT ME, DO IT ON DISCORD. MY USERNAME IS thebetternaen (Please don't contact me on Discord.)
