ai-mathematical-olympiad-progress-prize-2

[inland wedge]

Welcome to the 2nd AIMO progress prize. We have launched today!
I'm Frieder, the AIMO Prize Manager, and I'm here to answer questions (although I'll more often be found on the Kaggle forum - post there if you really need an answer quickly 😉).

Problems are harder this year (national Olympiad level) and AI-resistant (your favourite open-weight LLM will score close to zero).
The prize pot has doubled to two million dollars, and you are allowed to use the latest open-weight models.

Surprise us with some cool new math-AI models, and climb up the Kaggle leaderboard! 🪜

[lavish bear]

SO EXCITED

[shut mica]

I am getting

GatewayRuntimeError: (<GatewayRuntimeErrorType.SERVER_RAISED_EXCEPTION: 3>, "name 'submission' is not defined"), is anyone else getting the same error?

[shut mica]

https://www.kaggle.com/code/huikang/qwen2-5-math-1-5b-instruct

[outer sonnet]

I'm newbe in LLM and sorry if is a too newbe question but... is https://huggingface.co/AI-MO/NuminaMath-7B-TIR a "decoder-only" and "autoregressive" model? I'm tryng to replicate the text generation given by predefined pipe to understand the process and I'm struggling a bit

[sick grotto]

It's a finetune of a deepseek math model.

it is indeed decoder-only and autoregressive.

[outer sonnet]

Thank you very much

I've been able to roughly reproduce text generation by taking last logits of every iteration to predict next token (since mean over all of them was not working at all, I think because it is causal and each token doesn't sees any token after him, not sure about that). I've obtained a quite similar text, but not the same. Some one knows where I can find the model.generate.forward implementation or knows how it proceeds?

[acoustic rampart]

Is there any guide on how to get started with the solution?

[acoustic rampart]

So you’re saying the DeepSeek math model is free to use? Am I correct?
How can someone use and customize it for this competition? Is there a guide available for fine-tuning it for these mathematical problems?

[sick grotto]

Yeah. The competition allows the use of open-source models.

As for finetuning models in general, there are many guides on the internet. It is important you find good math datasets. Take a look at solutions for the previous edition of this contest for some inspiration. I’m sure there’s also datasets mentioned in the discussion.

[acoustic rampart]

I would greatly appreciate your help; I can't find the model or the guide on how to do this.

[sick grotto]

Search huggingface, deepseek 7b math

More explanation is given there

It’s also available on Kaggle by just searching it in the models tab. since then other good math models have also been released, such as Qwen2.5 Math series

[acoustic rampart]

@sick grotto ,The model seems costly. How can I access it for free on Kaggle?

[sick grotto]

I have no clue what you mean

Its a freely available model

What do you mean costly?

[acoustic rampart]

How would you run the model in Kaggle? we need APIs for that r?but It's showing prices on their site.

[sick grotto]

You can’t make API requests when making a submission to this contest, so you don’t have to worry about that.

The model can be downloaded

[north island]

yes

[sick grotto]

And can be used in kaggle by loading it with code in a notebook and running it

I recommend looking at other people’s public notebooks to see how they did that.

[acoustic rampart]

thanks @sick grotto

[outer sonnet]

Perhaps it does Beam Search?

[modest sun]

@inland wedge Is it possible to relax a rule about Googlers potentially receiving a prize?

[inland wedge]

@modest sun Googlers, as well as any other large lab, are very welcome to participate as long as they follow our rules, in particular or transparency requirements (meaning: all LLMs have to be open weight, you have to document your training process, etc.). 😉

[modest sun]

It's about this part of the rules:
I found out that Kaggle is Alphabet's subsidiary
B. Unless otherwise stated in the Specific Competition Rules above or prohibited by internal policies of the Competition Entities, employees, interns, contractors, officers and directors of Competition Entities may enter and participate in the Competition, but are not eligible to win any Prizes. Individuals or entities who were engaged, employed or contracted by the Competition Sponsor or its affiliates to advise on the Competition are prohibited from entering the Competition. "Competition Entities" means the Competition Sponsor, Kaggle Inc., and their respective parent companies, subsidiaries and affiliates. If you are such a participant from a Competition Entity, you are subject to all applicable internal policies of your employer with respect to your participation.

[inland wedge]

Ok, I see what the issue is, let me get back to you on that

[maiden flicker]

Hi !
So we are back! It's been a couple of days, and the competition is already eating my Kaggle GPU quota and my colab compute units 😅 , and I love it! 😎
Good luck everyone!

[dense vale]

excuse me kinda even newbie doubt but do we only ve option to use it via torch not tensorflow or kerasnlp or spaCy ? (cause i ve never tried pytorch yet)

[tight cobalt]

u should be able to use them all via import statements in your python code. And scikit, numpy, etc. as well. If it is a freely available and widely used python API, you should be good. I'd personally go with TensorFlow over Pytorch, but that's just me.

And Keras is a layer built on top of TensorFlow. btw. Think of it as boilerplate code that interfaces with it, (built by
François Cholletv for use with TensorFlow for AI tasks). You using it, you are using Tensorflow, in a nutshell.

[outer sonnet]

SpinDoctorWalker is right, HugginFace usually gives both choices. But in this specific model looks like only torch is available. I'm sure it can be translated, but it won't be a trivial task.

[tight cobalt]

@dense vale The only things you can't use are things such as chatGPT, or Gemini, or other LLMs u would access online

[dense vale]

u simply meant i will need to be good at pytorch first ??

[tight cobalt]

u can go with either framework (keras or pytorch). They are both for general AI exploration.

also, kinda expected you can import either in all kaggle competitions ;

[dense vale]

sir i meant with regard to the tuned deepseek model (numina one) can we access it with frameworks other than pytorch ?

[tight cobalt]

had to check it @dense vale . A definitive no, i think, because this requires you to access a resource that is online. And one of the requirements of the competition is no external internet access. So you can make of all the APIs you want, just not ones that interact with an external source, when a potential user interacts with it.

and it doesn't even count as a freely available source of information either (which is allowed in competition)

[dense vale]

btw i m confused how will we run the kaggle notebook(s) without external internet ?

[tight cobalt]

In your example, DeepSeek is an external service, so to speak. And they are telling you they don't want that as part of the solution they are seeking. Make sense?

[dense vale]

oh ok we can use the downloaded models though right ?

but still kaggle notebook how the code can be executed without internet ?

[tight cobalt]

yup. and i quote from competition page: "Freely & publicly available external data is allowed, including pre-trained models "

[tight cobalt]

It's not that u don't use the internet, or using the cloud, when submitting your algo, and running it. It's more about saying you develop a program that can run offline, once done. A program/ AI that can reason mathematically: of what use is an internet connection. It can run offline. Ideally.

[dense vale]

ok i get it now 😅 (but i m kinda new to llms just learnt the basics of nlp with spaCy could u recommend me what should i learn now kinda confused)

[inland wedge]

To clarify: Your model has to be open-weight (that you will probably download from somewhere), and it then needs on the Kaggle container without internet, within the given GPU budget. For AIMO1, a number of top-scoring teams actually spent significant amounts of time with engineering tricks, to get big LLMs to run on the given compute budget. That is one way to approach this competition, but not the only way ...

[dense vale]

Thanks sir but can you please also elaborate other ways ?

[thorn yacht]

Hi, can someone suggest a nice dataset that I could use for becnhmark or fine-tuning. is the dataset from the first competition available?

[acoustic rampart]

Yep same error

[analog sparrow]

The MCTS comp is eating up all my GPU quota haha, probably will only join after it ends 🤣

[inland wedge]

@dense vale It is all written in detail in the competition rules, in particular check out the section about using tools (https://www.kaggle.com/competitions/ai-mathematical-olympiad-progress-prize-2/rules)

[shut mica]

I published a notebook with the highest public score of 5 https://www.kaggle.com/code/huikang/qwen2-5-math-72b-instruct-with-tir

[dense vale]

does it perform well on aime 2024 problems ?

[acoustic rampart]

hi ,fine turn the model on large datasets from huggingface and other data source , it might help in improving the score

[dense vale]

anyone familiar with LEAN a programming language used by deepmind's ultra model claiming good score at IMO problems ?

[storm atlas]

is using two separate models or two lora adapters on the top of the same model in line with the rules ?

[solemn nymph]

could anyone explain how do i resolve this error? Thank you

/opt/conda/lib/python3.10/site-packages/torch/cuda/init.py:230: UserWarning:
NVIDIA L4 with CUDA capability sm_89 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_60 sm_70 sm_75 compute_70 compute_75.
If you want to use the NVIDIA L4 GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/

warnings.warn(

this error

[storm atlas]

afaik, it's for theorem proving

[storm atlas]

share your code, do you use vllm ?

[hoary granite]

yeah it's the mathematicians' favorite proof assistant, this book teaches it
https://hrmacbeth.github.io/math2001/

coq is also quite famous due to these nice books
https://softwarefoundations.cis.upenn.edu/

[dense vale]

Thanks alot for sharing these

[acoustic rampart]

share your code

[solemn nymph]

@acoustic rampart @storm atlas Here is the screenshot of the code that lead to this output

I went to the competition page, opened a new notebook from there and ran this code block, and then I get this error.

[shut mica]

probably need a different Pytorch installation? (no idea haven't done this)

[solemn nymph]

I tried it, when I make my own notebooks I get this error, even tried changing PyTorch and cuda versions

I also tried by copying someone else’s notebook from the competition and edited it; then it worked properly

[modest sun]

@inland wedge Another gentle ping about Googlers participation 😄

[inland wedge]

@modest sun Please give me a bit of time - the competition will run long enough, so there is no need to rush. We have an internal process, and I'm working through it, but there are several parties involved whose feedback I still need to collect before we can arrive at a decision.

[modest sun]

Ah yes, of course, sorry for that, I won't ping you anymore

[inland wedge]

It may indeed look like I've forgotten, since I haven't posted, but behind the scenes we're pretty active 🙂

I'll be sure to let you know when a decision has been reached

[acoustic rampart]

Check your CUDA and PyTorch versions—they need to be compatible.

[inner warren]

hello, i'm new here, i get one team

[shut mica]

lol i saw this ad (censored identifying information, I am not promoting this)

[analog sparrow]

Is this Numina ? Lol

But honestly it looks like a scam

[shut mica]

idk maybe if your annotations are bad you don't earn money

but educational to learn what are they annotating

[tawdry vale]

@modest sun You should have recieved an email from Maggie explaining Google's policies around this. I'll send you an internal gchat to continue this rather than talking on discord.

[thin vector]

Hello, I've just joined, are we allowed to install/use extra libraries in our notebook? I noticed that "Internet access disabled" is mentioned in the code requirements and I can't seem to pip install extra dependencies on my copy of the Demo submission

[outer sonnet]

upload them as datasets and install from them

[shut mica]

There is this utility scripts feature but I havent researched how to use it

[tight cobalt]

This limitation does seem a bit ridiculous

[fair shoal]

Hi guys, I'm new to the competition and I'm having a great deal of trouble with vllm. Currently, I am installing vllm with the Kaggle package manager, but I always get errors when importing vllm, or when trying to start an LLM API server with vllm. The error is usually that vllm tried to call some pytorch or torchvision method that does not exist, but the exact error depends on the exact versions of pytorch, torchvision and vllm that are installed, I've tried dozens of configurations now an am still having no luck. Any suggestions to get vllm working would be greatly appreciated

[minor inlet]

Hey!! I encountered the same issue with vllm on Kaggle, and after some searching in the forums, I found that it's a common problem due to version mismatches between libraries used to run vllm.

Here’s the solution!! Coming from @pseudo remnant
Just add a Utility Script that fixes these issues. Go to your Kaggle notebook, select Add Input > Utility Script, and search for "vLLM Installation Fix" or abdullahmeda/vllm-installation-fix. Add it to your notebook, and after that, you should be able to import vllm without issues. Easy fix!

If you’re new to adding inputs or need help with it, feel free to ask!

[fair shoal]

Thank you so much, I wasn't expecting such a quick and easy fix

[north wyvern]

guys, im looking for a team for this one, im into math and i need some exp in ml

and want to chill on call while working on it

[lime raptor]

I'm doing BS IT with expertise and interest in machine learning previously I have joined two competition, if you would like to connect with me then dm

[shut mica]

wrote something https://www.kaggle.com/competitions/ai-mathematical-olympiad-progress-prize-2/discussion/546772

[dense vale]

anyone in fimo ??

[austere pecan]

Hi,
Is there a channel to discuss and understand solutions of the problems?

I tried solving the "3 Airline companies" Problem by hand. In my solution, I found the greatest consecutive days could be 99 days. I checked the solution to find the correct answer is: 79 days.
For my solution, the A120 flight departs along with first A100 flight. I have attached the image to this thread.

Could someone comment: is my solution a valid alternate solution or am I violating some pre-requisite condition?

[shut mica]

The question text here

Three airline companies operate flights from Dodola island. Each company has a different schedule of departures. The first company departs every 100 days, the second every 120 days and the third every 150 days. What is the greatest positive integer d for which it is true that there will be d consecutive days without a flight from Dodola island, regardless of the departure times of the various airlines?

If I find a configuration of departure times where the the maximum interval is less than 80, the answer cannot be 80 or above. The "original solution" has shown such a configuration.

The question asks for a minimum possible "x" over all possible configurations. Given a configuration, "x" is the maximum interval.

The question did NOT ask for a maximum possible "x" over all possible configurations.

[stiff hill]

that airline question is interesting, because I looked at it and guessed the answer was 79 or 80 (didn't have time to reason which) in < 10 seconds without doing any calculation. So it's not true that these new problems in progress prize 2 are all very difficult to guess

[outer sonnet]

They're exactly 1/100 hard to guess

[shut mica]

20/50 gets the prize. So close!

[lethal magnet]

Hi, I'm new in kaggle, what might be the reason of I can't submit?

[runic path]

Click save version first

Save/run all

[dim jetty]

any one tried to use RAG?

[wispy grail]

hi

[pseudo willow]

Can someone clarify if it is ok to scrape data from https://artofproblemsolving.com/wiki/index.php/AIME_Problems_and_Solutions and use it for fine-tuning and validation? Are there any license issues?

[tight cobalt]

I don't think you can scrape data directly, during run of program. You can scrape though and create a result of that scrape saved as data to your Kaggle account (aka, an extarnal model)

[hot yoke]

Is it possible to see the running time of submission?

as in this contest it might be pretty important

[shut mica]

I investigate this question in https://www.kaggle.com/competitions/ai-mathematical-olympiad-progress-prize-2/discussion/549184

[worthy kraken]

Hi,

I wanted to discuss the professional aspect of fine-tuning approaches.

To clarify, I have no issue with teams fine-tuning models to achieve higher scores. However, it seems that very few participants actually have access to sufficiently powerful GPUs for such tasks.

This brings me to wonder:
What are the actual limitations of fine-tuning these models?

Apologies if I’ve missed any prior discussions on this topic. If there are relevant discussions, please feel free to point me in their direction.

[lime raptor]

hello @everyone I am little bit confused about dataset and other files that are showing in side bar, i want to know how start with this problem, where can i find dataset, here are two one is reference showing only test showing 3-5 problem, test set showing i think approximately 10 problem and one pdf contain 10 problem so can anyone guide me about how to start working on this from where can i get dataset and more. kindly help me thanks.

[minor inlet]

Anyone still interested in teaming up?

[lime raptor]

yes I am

[tardy dragon]

Is there a vacancy?

[fluid maple]

I have the same issue!!! Can someone help??

[lime raptor]

I think use prompts

[cursive wolf]

Is internet allowed during submission?

[minor inlet]

No It is not

[limpid ether]

Can anyone explain the message "Your submission file must be named submission.parquet" shown when you try to submit your notebook? I don't see anything about saving a parquet file in the provided example submission notebook.

[cursive wolf]

is gemma 2 allowed?

[green flint]

Would anybody be willing to teamup for the AI Mathemtical Olympiad competition? If so, respond

[compact vault]

I'm willing to team up

[shut mica]

I want to write a Kaggle post detailing the public information (i.e. where they worked, what did they recently published) of the team members of NemoSkills and PolyMath, I wonder if it is appropriate

[fluid maple]

Can you please elaborate more on how to use prompts! Thank you in advance

[olive lark]

Hey I am begginer here.i don't know anything about kaggle.can anyone explain me the link between ai and mathematics is the exam over

[fresh prawn]

onebelowall1218 has been warned

Reason: Bad word usage

onebelowall1218 has been banned

Reason: Too many infractions

[frank wagon]

is anybody seeing this?

@shut mica I would like your thoughts on this

[shut mica]

Well they are from Nvidia they have GPUs

[grand palm]

Anyboyd from NVIDIA want to team up with a poor man and bring him to his first gold medal ?

[shell cloak]

hi guys my first ever submission on kaggle, is this normal running time ?

[tardy dragon]

Yes, it takes almost 5 hrs

[sharp loom]

😆 that is very normal. The first time I ran a heavy model it took me by surprise too....

what do you think would happen if the private test set has USAMO level problems. Would NemoSkills model hold or would the leaderboard be reversed??

[dense vale]

Why no one tried rStar approach yet ?

Is symbolic reasoning(Sympy) explicitly needed in rStar maths method ?

[mortal vault]

tried it but don't have the kind of compute available to actually do it well unfortunately 😔

[dense vale]

what do u mean ? does it require that much compute even for a single run ?

[mortal vault]

yeah if you see in the repo the first step is to generate a bunch of step wise training data using rollouts. i was running it using the 4xL4 setup and timed out after 5-6 hours and only got to like 5 rollouts

i tried smaller and smaller models, decreasing the amount of exploration in the params, didn't really make it close to being feasible

[dense vale]

wdym by 4xL4 setup btw 🤔 😅

[mortal vault]

um the 4 L4s provided for this competition?

[dense vale]

what L4s u mean ?

[mortal vault]

these??

[dense vale]

oh u meant that 😅

[dense vale]

what models did u try btw ?

[mortal vault]

i think i tried deepseek 1.5B, qwen 1.5B and 0.5B

[dense vale]

oh did u try phi4 ?

i was thinking to try 7Bs 😅

[mortal vault]

yeah my reasoning for only trying those models was that if i couldn't get the rollouts to finish in time using the small models, no way i could with the larger models

[dense vale]

make sense so does that mean rStar isnt feasible at all for this 😞

[mortal vault]

honestly, unfortunately almost any type of method that requires even a little bit of finetuning is out of reach for anyone without access to external GPU compute

[dense vale]

sensible 😔

[dense vale]

anyways did u try mathstral ?

having problem loading it smh or im not getting how to use it with transformers

[mortal vault]

nope i haven't

[dense vale]

u r at 1334th in lb btw ?

[mortal vault]

i have no idea really, most of my work has been unsubmitted. i just submitted the early submission awq qwen 32B model for benchmark and left it at that

would've liked to have done more with this competition, but too busy this month unfortunately

[dense vale]

with jane street one ?

[steady stirrup]

Our team did not look into the problem yet, but the public nbs give a pretty good score tbh

but

apparently they are pretty unstable

[dense vale]

nbs ???

[steady stirrup]

notebooks

I am in 2 other comps concurrently

hence I sheldom have gpu remaining at the end of the week

• 4xL4 worker groups consume twice as much quota(iirc)

[dense vale]

||can u teach me rStar method sir||

[dense vale]

||anyways lemme get u dqd using it as evidence that u use alt accounts n instead of the person u replied is evident that it was ur alt acc||

[steady stirrup]

😂

[steady stirrup]

😂

[steady stirrup]

• I am extremely busy with non AI stuff, which I prob should be doing more than AI rn 😂

[dense vale]

🛐

[mortal vault]

nah college stuff

jane street i barely gave any time to too

[lost seal]

how much score did this get

[mortal vault]

i didn't submit it because i didn't have compute to actually finish the whole process

[lost seal]

also can you just train on whitelisted models and that would be cosidered okay to use right

[lost seal]

it scores whatever it has answered right?

im not sure how the entire evaluation part works

but isnt that why smaller models are getting better scores than bigger ones like 32b

since theyre answering mmore questions

[mortal vault]

i didn't finish generating the stepwise training data required for the SFT for rstar math bruh 💀

[lost seal]

what

i dont think you have to train and submit in the same notebook

also rstar then sft wth

sft is would just bad for this

[mortal vault]

sft is part of rstar...

please look at the steps to implement rstar

i think a lot of people misunderstand how rstar works because of the recent RL hype

[lost seal]

oh like coldstart training

[mortal vault]

not really no

[mortal vault]

posted some of my unexplored ideas here
https://www.kaggle.com/competitions/ai-mathematical-olympiad-progress-prize-2/discussion/567419

[dense vale]

Is sft really part of it ?

[dense vale]

Oh right thnx

[steady stirrup]

I did not read the paper, but in their abstract they specifically mentioned without finetuning right?

@mortal vault

[mortal vault]

oh dear

um

so i meant rstar math when i was saying rstar

[steady stirrup]

😮

nvm

[mortal vault]

just check out the repo yall https://github.com/microsoft/rStar
and read through this discussion https://github.com/microsoft/rStar/issues/6 it helped me understand how the thing works

[steady stirrup]

thanks

[steady stirrup]

I get that I am an idiot, why are you making fun of me?

[stiff pelican]

Is it too late to join competition?

[dense vale]

Its never too late to join

Each time in a featured competition top teams in lb we see newbies

Who 've never been into any previous competition

[stiff pelican]

Oh, that's cool

Thanks!

[slow sonnet]

I'm new to this competition. I'm a bit confused with whitelisting of models. Do we need to use them as they are for the competition or can we finetune them with any public data offline, use the finetuned models for competition and publish them later?

[dense vale]

@steady stirrup

[steady stirrup]

im not sure, im not much into this comp, but I believe you can finetune

[dense vale]

But that whitelisting thing ??

[steady stirrup]

read the discussion page once

[dense vale]

I've read but still in doubt do we 've to request for whitelisting our fine tuned models to use them ?

[alpine coral]

anyone having issues saving and running notebooks? when i do that, not quick save, my notebook get queued for more than an hour. never seems to start...

[dense vale]

That's not normal ig

[alpine coral]

Yeah , never experienced it before on kaggle

[dense vale]

Could be a possibility u r trying to load a larger llm consuming lot of vram than usual

[alpine coral]

it hasnt gotten to the stage where the model is loaded yet, it hasnt even started running the notebook. the activity log says its queued

and its just the normal 32B preview in 4bit awq

[steady stirrup]

It happened to me once

it might be because of high demand of l4 worker groups

im not sure tho

[fresh prawn]

ren_truecon has been warned

Reason: Bad word usage

ren_truecon has been banned

Reason: Too many infractions

[slow sonnet]

Thanks.

[dense vale]

Thnx alot sir

[stiff pelican]

I have a question. When you guys train an LLM model, how do you do it? Training the entire model takes a very long time. My model is deepseek-r1-distill-qwen-7b-awq-casperhansen which is in public notebook.

To many time, i made dataset with 1/10. However, still long time with l4 x 4 gpus.

[dense vale]

Don't u use bitsandbytes ?

If yes also try gradient accumulation

If yes also try instruction tuned slms around 135-200M ones

[stiff pelican]

Nope, i try my own code to train.

Then, i try to change my code to using Lora or Quantization...

I'm not very experienced with LLMs, so there might be some mistakes in my approach.
Theoretically, I've read the R1 paper and implemented the RL method from that paper in my own way, and currently, I'm training the model based on it. I'm wondering if applying LoRA would be an appropriate approach in this case. Could you advise me on this?

[dense vale]

Most ppl said they were too expensive so I'm not considering I'm considering anyways ReFt mostly though yet learning it firstly not yet tried but it seems good enough

[stiff pelican]

Thanks for the suggestion! Actually, I haven't looked deeply into ReFt yet, but based on your advice, it seems like a promising approach. I'll check it out and see if it fits better than LoRA for my use case. If you happen to try it before me, let me know how it goes!

[cursive wolf]

Hey I have a question...so after my GRPO run I saved lora weights and did inference using vllm + lora weights...is this why the model was not as good...should i have saved the entire model instead?

[stiff pelican]

From what I've seen in the discussion, there seems to be a tendency for trained models to perform worse. Please check the following link:
https://www.kaggle.com/competitions/ai-mathematical-olympiad-progress-prize-2/discussion/568061

[cursive wolf]

thank you this is very hepful

[stiff pelican]

But know it is changed...
https://www.kaggle.com/competitions/ai-mathematical-olympiad-progress-prize-2/discussion/568509#3151143

[cursive wolf]

oh lord

[distant fulcrum]

excuse me, is fine-tuned version of models released after October 1, 2024 allowed in this contest?

I just noticed this in the rule.

[shrewd hemlock]

Machine Learning Algorithms You Never Knew Existed, But Are Quite Useful https://medium.com/pythoneers/machine-m. D

[past grotto]

Hi everyone, our team is new and would appreciate any guidance. Our submissions consistently fail with an "unhandled error at runtime," but we notice the Logs tab always shows "successfully ran in x seconds." We're curious if the run in the Logs is separate from the actual scoring run. Is it common for the notebook to run twice? Thanks in advance!

[frank eagle]

Just a quick question, if we submitted our notebook at the last minute before the deadline, and it was running the pubilc test, will the result still count?

[dense vale]

Nope no point better luck/try next time

[warm fossil]

Is it possible to modify the notebook selection at this point?

[dense vale]

Wdym ?

[warm fossil]

Kaggle auto-selected two submissions. We would like to replace one of the auto-selected submissions with a more recent notebook that has the same public score, if it is possible.

[dense vale]

If it's already over ofc not possible

But maybe as practice competition u might submit

[alpine coral]

will there be a aimo - progress prize 3?

[dense vale]

Next year

[frank eagle]

Will you publish the hidden dataset at some point,
or provide some ways for us to test our model for research?

[distant fulcrum]

What's the level of the third contest?National team selection or IMO?

[dense vale]

Yes in between