PARTICIPATION
Teams may enter either track or both. Use one system for all text tasks (T1–T4) and one for the vision task (V1). The same system may serve both tracks, but each track requires a separate submission and is evaluated separately.
Organizers will not provide a training dataset. Synthetic data, self-created data, and data collected from the web may be used.
Text / LLM models use four bands: Tiny (P ≤ 1B), Small (1B < P ≤ 8B), Medium (8B < P ≤ 30B), and Large (P > 30B). Vision / VLM models use three bands: Small (P ≤ 8B), Medium (8B < P ≤ 30B), and Large (P > 30B). P is the total number of model parameters. For mixture-of-experts (MoE) models, all expert parameters count, including experts that are not active for a given input. For vision-language models, the vision encoder parameters are included in P. Boundary values belong to the lower band. Each model-size band has a separate ranking within each track.
Any model architecture is allowed if it meets the parameter limits of its band and is compatible with the Hugging Face Transformers library.
The model must load and generate outputs locally with the Hugging Face Transformers library. Include model weights, required tokenizer and configuration files, any required processor files, and run instructions.
During evaluation, all outputs must be generated by the submitted model running locally through the Hugging Face Transformers library. External APIs, remotely hosted models, and online services may not be used to generate outputs.
A validation dataset will be shared for model development and hyperparameter selection. Final evaluation uses a private test set that will not be shared with participants; official rankings are based on those results.
Platform details, technical submission requirements, and the submission schedule will be announced separately.