Description: This stage utilizes the Honor of Kings 3v3 Changping Map. In addition to the main elongated lane, the map includes an upper and a lower jungle area. At each end of the map are the revival points for each team's agents, with the team's crystal located in front of the revival point. The crystal continuously produces minions for its team, which automatically advance towards the enemy's camp, attacking enemy towers, crystals, and heroes along the way. In front of the crystal is the team's defense towers, capable of attacking enemy heroes and minions entering its range. The jungle area is a primary source of economic and experience gains for jungler heroes. Players from both teams share the entire jungle area, although some entrances to the jungle are team-restricted. Agents can freely move and use skills within the map. Over time, agents gain gold and experience, with significant rewards for killing jungle monsters, enemy minions, enemy heroes, or destroying enemy towers. The team that successfully destroys the opponent's crystal wins the match.

Task Objective: Complete the adaptation and end-to-end deployment of the baseline model within the specified time. On the basis of ensuring the availability of the deployment model meets the standards, optimize the various inference performance indicators of the model running on designated iQOO mobile devices. Specific indicators are detailed in the following text.

Evaluation method: Conduct battle testing and performance testing on deployment models (divided into Linux battle packages and mobile deployment packages).

● Battle Test： The Linux battle package submitted by the participating teams and the reference model provided by Tencent AI Arena platform are used for multi round battles. In each round of the match, both models compete according to the test lineup combination and match each other's lineup combination one by one. Each combination switches sides and plays 2 rounds. After the battle is completed, calculate the victory rate based on the number of wins/total number of matches submitted in the model.

The participating teams need to ensure that the calculation results of the same test input processed by the submitted Linux battle package have an error smaller than a certain threshold compared to the mobile version. Otherwise, the submitted model will be deemed invalid and the test will fail.

● Performance testing：The platform regularly conducts batch testing on the mobile deployment packages submitted by each participating team, collects running results and performance data through multiple runs, and serves as a scoring basis.

Evaluation indicators：
A usability standard is set for the deployment models submitted by participating teams. Once this standard is met, performance scores can be calculated.

●  Usability standard
    ○   Baseline capability inheritance: Complete battle testing, deploy models against reference models, with a win rate of no less than 40%
    ○   Stability: Completed performance testing, deployed the model to run stably multiple times, without memory exceptions or crashes
●  Mobile inference performance
    ○  Calculate the comprehensive performance score based on the data collected in the performance test
    ○  Scoring sub-items (weighted by a certain proportion)
      ■  Time consumption - The mobile inference time consumption of the deployed model
      ■  Size - The mobile storage overhead of the deployed model
      ■  Memory - The device memory overhead during deployment model runtime
      ■  Power consumption - The device power consumption level during deployment model runtime