Обратите внимание, что скачивание данных с платформы выключено. Это сделано для того, чтобы вы решали задачу из среды ноутбуков, генерируя csv с прогнозами и отправляя его прямо в среде на Stack. Таким образом нам (организаторам) будет доступен код ваших решений и мы сможем проверить решения на соответствие правилам, описанным ниже.

You will need to train a machine learning model capable of classifying an animal's breed based solely on its text-based ASCII representation. Return the knowledge of breeds to humanity using only the power of the console and AI.

Task Limitations

• External Data: The use of any external datasets is strictly prohibited. You must rely solely on the provided competition data.
• Architectures vs. Weights: You are free to use any model architecture you want. However, there are strict rules regarding pre-trained weights:
  • Allowed Pre-trained Models: You may use pre-trained weights for only the following two models:
    • google-bert/bert-base-uncasedExample of loading and using the model:

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-uncased")
model = AutoModelForSequenceClassification.from_pretrained("google-bert/bert-base-uncased", num_labels=2)

text = "Sample text for inference"
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)
    logits = outputs.logits
print("Logits shape:", logits.shape)

• gpt2Example of loading and using the model:

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("gpt2")
tokenizer.pad_token = tokenizer.eos_token

model = AutoModelForSequenceClassification.from_pretrained("gpt2", num_labels=2)
model.config.pad_token_id = model.config.eos_token_id

text = "Sample text for inference"
inputs = tokenizer(text, return_tensors="pt", padding=True)

with torch.no_grad():
    outputs = model(**inputs)
    logits = outputs.logits

print("Logits shape:", logits.shape)

• Training from Scratch: For all other algorithms, models, and architectures (e.g., CatBoost, TF-IDF, IsotonicRegression, GigaChat3.1-702B-A36B, custom neural networks, and any predefined architectures), pre-trained weights and checkpoints are strictly forbidden. These models must be initialized randomly and trained entirely from scratch using only the provided dataset.

Description

The year is 2056. The era of graphical user interfaces is gone forever. Following the total domination of Claude Code, Gemini CLI, Codex, OpenCode, and other console-based AI agents, humanity has completely forgotten how to render pixels. Windows, buttons, and 4K monitors have been left to gather dust in museums - now, all of human digital life, from work to entertainment, flows exclusively through the terminal.

But one thing remains unchanged: the fundamental human need to look at cats and dogs.

To adapt old photo archives to this new reality, images of our furry companions were run through console converters and turned into ASCII art. Unfortunately, during the 'Great Console Transition,' almost all metadata was lost. We are left with thousands of text-based pet portraits, but we've forgotten which is a pug, which is a Maine Coon, and which is a Shiba Inu.

Example of an ASCII art image (CatBoost logo):

■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■
■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■
■■■■■■■———————————————————————\■■■■■■■■■■■■■■■
■■■■■■■||—————————————————————\\■■■■■■■■■■■■■■
■■■■■■■||OOOOOOOOOOOO/———\OOOO\\\\■■■■■■■■■■■■
■■■■■■■||OOOOO//——\O//———\\\OOOO\\\■■■■■■■■■■■
■■■■■■■||OOOOO//——\\//    \\OOOOO\\\■■■■■■■■■■
■■■■■■■||OO/—\\\——//\\.  ///—OOOOO\\\\■■■■■■■■
■■■■■■■||O/■■■\O—\\O\O//—/\■■\\OOOOO\\\■■■■■■■
■■■■■■■||O\\——\\\/\——//////——/OOOOOOO\|■■■■■■■
■■■■■■■||OOOOOO\//—.— \\\/OOOOOOOOOO///■■■■■■■
■■■■■■■||OOOOOO//       \\OOOOOOOO///■■■■■■■■■
■■■■■■■||OOOOOO\\       //OOOOOOO///■■■■■■■■■■
■■■■■■■||OOOOOO\\\————////OOOOOO///■■■■■■■■■■■
■■■■■■■||OOOOOOO\\—————/OOOOOO////■■■■■■■■■■■■
■■■■■■■||—————————————————————//■■■■■■■■■■■■■■
■■■■■■■———————————————————————/■■■■■■■■■■■■■■■
■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■
■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■

Note: This image is not from the dataset; it is simply an example of what an ASCII art image might look like. Your models will need to work with the actual data from the dataset.

Data

Each object in the dataset is a former photograph converted into a string of 128 by 128 characters (with \n separating the rows). Characters of varying density convey the lighting, shadows, and contours of the animals.

• train: A set of ASCII images with known breed labels for training your model.
• test: A set of ASCII images for which you need to predict the animal's breed.

It is highly recommended to check out the provided baseline notebook to get a better understanding of the data.

Evaluation

Participants' predictions will be evaluated using the F1-micro metric.

Submission File

    id,label
    2,0
    5,0
    6,0
    etc.