Checklist HAIO 2025 Summer National Final · Task 4
Epic Emojis
English title: Epikus Emojik
Fine-tune the T5-based EmojiLM model to translate English film titles into emoji sequences and analyse its cross-attention.
The task
Emoji Emma wants every film title summarised as a short emoji sequence. The contestant fine-tunes a pretrained text-to-emoji model on a small dataset of film titles and emoji sequences (the one-pager states it covers the 1,000 top-rated IMDb films).
Task 1 (10 points): show 10 random examples and count unique titles (after removing duplicate rows) and unique individual emoji characters. Task 2 (5 points): drop rows with missing values and duplicates, then sample 60 random examples for training. Task 3 (5 points): load the tokenizer and model KomeijiForce/t5-base-emojilm (inputs use the prefix "translate into emojis:"). Task 4 (20 points): prepare the data as a Dataset or DataLoader with padded inputs and label padding set to -100. Task 5 (10 points): write a generation helper and a BLEU evaluation over the training set for the untuned model.
Task 6 (25 points): fine-tune with early stopping based on the training loss (no validation set), recording the loss per epoch. Task 7 (3 points): print title, generated and reference emojis for the first 10 titles. Task 8 (7 points): compare with the untuned model, plot the training loss and compute BLEU after training. Task 9 (15 points): plot at least 5 heatmaps of decoder–encoder cross-attention weights for at least 5 examples.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
In English
Some of this task's files were published only in Hungarian. SOTA translated that file into English on 16 September 2026.
Read the task description (one-pager) in English
Epic Emojis
English translation by SOTA – AI Community of the Hungarian original. Licensed CC BY-NC-SA 4.0, like the original. Organisers who would like this translation removed can email [email protected].
Magyar MI Diákolimpia (Hungarian AI Olympiad)
Epic Emojis 😎 🎉 💬 🔥 (Epikus Emojik)
Task description
24 May 2025
1. Epic Emojis
Emoji Emőke is a self-appointed film aesthete who is obsessed with the world of cinema and with emojis. She has devoted her life to summarising every existing film title with a single, well-chosen emoji sequence (4–5 emojis).
Her philosophy of life:
"If a film cannot be described with three emojis, it does not even deserve to be watched by me!"
It takes Emőke a lot of time to come up with the emojis for every film title, so she has no time left for what she really loves doing: preparing for the AI olympiad.
Your task: help Emőke automate the translation of films into the language of emojis. You will do all this with the help of a pretrained model and a database, which you will receive. The various subtasks introduce you to the training cycle of language models, and at the end Emőke will have a model at her disposal that will greatly help her in her work. You can solve the subtasks in a Jupyter Notebook (Google Colab is recommended; a local environment is also accepted, but training is slow there).
Examples:
- Lion King 🦁 👑
- The Godfather 🤵 🔫 💰 👪
- The Lord of the Rings: The Two Towers 🧙 💍 ⚔️ 🌲 🏰
2. Scoring
- Data exploration – [10 points]
- Data cleaning – [5 points]
- Loading the model – [5 points]
- Data preprocessing – [20 points]
- Baseline evaluation, model without fine-tuning – [10 points]
- Training the model – [25 points]
- Evaluation on the training data – [3 points]
- Visualisation and performance comparison – [7 points]
- Visualisation of the attention mechanism – [15 points]
Maximum score: 100 points
3. Technical information
The Google Colab interface is recommended for the solution because it provides a graphics card, with which training is about 60 times faster than on a CPU. The task can also be solved locally, but reaching the maximum score by training on a CPU is almost impossible within the time available for solving the task.
The task requires a model available on the Hugging Face platform that was pretrained on a text–emoji dataset, but not on film titles. This is important, because it means that we can fine-tune the model for our own purposes on a small dataset. All the necessary information and the required libraries can be found at the beginning of the notebook.
The dataset contains the 1000 top-rated films on IMDB and is available at the following link:
🔗 Film title – Emoji dataset
After solving the task, regardless of whether you worked in Google Colab or locally, you must upload the .ipynb file to the CMS system for the corresponding task. No other file needs to be uploaded.
4. Useful resources
Translated by SOTA. The Hungarian original is the official version and wins wherever the two differ. Original by the Hungarian AI Olympiad (ELTE Faculty of Informatics), licensed CC BY-NC-SA 4.0; this is a translation of the task one-pager, and the official English notebook of the task is linked on this page. This translation is shared under CC BY-NC-SA 4.0, the licence of the original. If you organise this olympiad and would like the translation removed, email [email protected] and we will take it down.
At a glance
- You get
emoji_dataset.csvwith columns title and emojis (downloaded with gdown; also in the repository folder adatok/emoji-epics) and the Hugging Face model KomeijiForce/t5-base-emojilm.- You submit
- The completed .ipynb notebook uploaded to the CMS.
- Scoring
- 100 points (10 + 5 + 5 + 20 + 10 + 25 + 3 + 7 + 15). Task 6 is scored by the final training loss: below 1 → 25 points, below 2 → 20, below 4 → 15, below 5 → 10.
- Rules
- A GPU (Google Colab) is recommended; the one-pager notes that training is about 60 times faster than on CPU.
- Work only on the provided lab machines (Windows); own laptops are not allowed. Google Colab is recommended; a local Python environment is also provided.
- At most one NVIDIA T4 GPU; stronger GPUs (e.g. A100, V100) lead to disqualification.
- General internet access, but code-completion and LLM services (e.g. GitHub Copilot, ChatGPT, Claude) are forbidden; the free tier of Gemini 2.5 Flash integrated in the Google ecosystem is the only exception.
- Publicly available documentation, articles and books may be used during the practical part; audio-visual material is forbidden; communication is forbidden.
- Mandatory screen recording with OBS Studio for the whole contest.
- Format
- Summer National Final (Nyári Országos Döntő), 24 May 2025, on site at ELTE, Budapest. Second (practical) part: two hours of programming after a 15-minute briefing; four 100-point tasks (CV, ML, NLP, RL). Solutions (.ipynb) are uploaded to the CMS.