No provided datasets
Data engineering is part of the challenge. You are responsible for sourcing, scraping, or generating your own dataset to train and validate your model.
Robust evaluation
You must design and implement a proper evaluation pipeline — just like in real-world engineering. Metrics matter as much as model quality.
Hidden test set
Optimising for the public leaderboard alone is not enough. Final solutions may be stress-tested against a hidden dataset to ensure true generalisability.
Monitor LLM safety by tapping into internal representations. You will implement a latent probing pipeline using Gemma to detect toxic or harmful user requests — extracting embeddings across layers and training linear classifiers to identify where safety-critical features are encoded. A computationally efficient alternative to full guardrail inference.
How to get started on Codabench
- 1Click the competition link above and sign in or create an account using the same email you registered with the cohort.
- 2Go to My Submissions and download the starter kit from the Files section.
- 3Return to My Submissions and submit your starter kit.
- 4Refresh and go to Results — you should appear on the leaderboard.
Mental health discourse is full of subtle cues that standard sentiment tools miss. Using Gemma's internal embeddings across different layers, you will train lightweight linear classifiers to decode psychological states and distress signals — building a computationally efficient framework for non-invasive emotional monitoring.
How to get started on Codabench
- 1Click the competition link above and sign in or create an account using the same email you registered with the cohort.
- 2Go to My Submissions and download the starter kit from the Files section.
- 3Return to My Submissions and submit your starter kit.
- 4Refresh and go to Results — you should appear on the leaderboard.
Standard tokenizers over-fragment African words into meaningless pieces, inflating compute costs and hurting performance. You will build a localized tokenization pipeline trained on native African corpora, then implement SuperBPE — a hierarchical "superword" merging strategy that preserves semantic units across morphologically rich language families.
How to get started on Codabench
- 1Click the competition link above and sign in or create an account using the same email you registered with the cohort.
- 2Go to My Submissions and download the starter kit from the Files section.
- 3Return to My Submissions and submit your starter kit.
- 4Refresh and go to Results — you should appear on the leaderboard.
Build a Small Language Model for legal advice and regulatory compliance. You will source and curate data from statutes, case law, and legal documents, then train and refine a model capable of addressing legal queries. Flexibility on architecture, fine-tuning approach, and alignment strategy (including RLHF) is encouraged.
Train a domain-specific Small Language Model to address real-world challenges in agriculture and climate. You will source and curate a high-quality dataset reflecting meaningful agricultural and climate contexts, then design, train, and refine a model that produces coherent and contextually relevant outputs.
Build a safety-critical SLM for medical advice and clinical decision support. You will curate data from peer-reviewed journals, clinical guidelines, and anonymized healthcare resources, with strong emphasis on data accuracy, patient privacy, and medical ethics. Guardrail implementation — including fact-checking layers or RAG integration — is strongly encouraged.
Preserve and celebrate African storytelling by training an SLM to generate traditional folktales. You will curate data from oral histories, translated anthropological texts, and published African literature. The challenge is capturing cultural nuance, dialect representation, and authentic oral narrative structure — not just fluent text generation.
Customer complaint systems are overwhelmed by unstructured text. You will build an end-to-end automated complaint classification pipeline using a transformer architecture (BERT or a distilled variant) — handling colloquial language, spelling errors, and class imbalance, then evaluating with precision, recall, and F1 across multi-class categories.
Build the retrieval engine at the heart of a RAG system for smallholder farmers. Given a farmer's question, return the most relevant documents from a knowledge base covering crop diseases, soil management, and climate adaptation. You are scored on nDCG@5 — beat the TF-IDF baseline with BM25, dense embeddings, rerankers, or hybrid approaches. No GPU required.