Current Competitions

Subnet 1 Apex competitions registry

1. Matrix Compression

The first competition - Matrix Compression - explores how small neural activations - both forward and backward - can be compressed while still retaining all their original information. Reducing activation size enables faster data transfer across the internet, a crucial step toward making distributed training more efficient, as it’s often constrained by network bandwidth. The top-performing algorithms from this competition will be integrated to enhance training on subnet 9 IOTA.

Evaluation

Miners aim to optimize the following:

  • Compression Ratio - How small the compressed solution is on disk versus the starting matrix.

  • Time - How fast the compression/decompression algorithm runs.

To surpass the current winner of the competition, a miner must earn a score of at least 1% higher than the current top score. If there is no current winner, then a miner must earn a score of at least 1% higher than the baseline score.

  • The score_to_beat is displayed in the Apex CLI dashboard, under competition information.

Score is calculated by:

score = np.clip((1 - compression) * (1 - task_time / (1 + 0.012)), 0.0, 1.0)

  • Where task_time includes both compression and decompression.

  • Compression is calculated by compressed_file_size / original_file_size.

  • In the compression of matrix mechanism, Similarity means that after you compress and then decompress a matrix the resulting matrix must be identical to the original - no changes in any values. Tasks performed with Similarity < 1 are not rewarded.

Other Matrix Compression settings

  • The submitted code remains hidden for 24 hours, after that it is open-sourced to the community to ensure fast innovation cycle.

  • The round time is 2 days. This means that the task pool is renewed and the logs for the competition are published every two days.

  • The emission Burn Rate currently is 90%.

  • "The winner takes it all" format means that the winning miner receives all the remaining 10% of miners emission as a first reward. If the same solution continues to stay on top, each subsequent reward for that solution will decrease linearly over 10 days, in accordance with the emission burning mechanism.

  • Each round, 30 matrices are randomly selected from a large matrix pool and used to evaluate all miners’ submissions. In the next round, a different set of 30 matrices is selected and used for evaluation again.

For Miners

View the matrix compression baseline miner solution provided as an example.

The manifest.json file contains a sample of 1000 matrices of a large matrix pool, including both matrices that have already been used for evaluation and matrices that have not been evaluated.

Note: The R2 bucket does not list files directly. Use the manifest to find and download samples.

  1. Download the manifest file. This file contains the full list of available sample filenames.

  2. Open the manifest and copy the name of any file you want to download. For example: example_001.txt.

  3. In your browser (or with curl/wget), go to: https://pub-77097c3387c340de9ff1bd5e5b443d8d.r2.dev/<file_name> replacing <file_name> with the name you copied from the manifest. For example: https://pub-77097c3387c340de9ff1bd5e5b443d8d.r2.dev/example_001.txt.

Then, continue to the Apex CLI guide to submit a solution.

PreviousIncentive MechanismNextMining

Last updated