Gitcoin RPG notes

tags: gitcoin Notes

Updated by June 2021

Goals

• To perform a dry-run of the technical anti-sybil workstream during round
• To make it fun and educative to manage the anti-sybil work

Roles

• Contribution data generator: Danilo
  • Sub-roles: dishonest contribution generator & honest contribution generator
• Machine Learning operator: Jesse Tao
• Human Evaluator: Jiajia
  • Responsible for manually flagging select labels that were predicted from the lastest run
• MC: Zargham
  • Subrole: Final report generator

Rules of the game

• Win conditions
  • For dishonest generators: funelling money away through the sybil tax
  • For honest generators:

Procedures

Bootstraping cycle

Flagging Cycles (iterative rounds)

1. Data generators create additional rows and provide metadata (labels) to the MC
2. Machine Learning Operator trains the subset-supervised ML algorithm and uses it to extrapolate over the all the existing dataset
   • (possible to change heuristics over time)
   • The result of the extrapolation is a tabular data with users and ids and containing the model results (eg: label and confidence)
3. Machine Learning Operator selects a subset of the extrapolation results and hand over to the human evaluator
   • (eg: the subset could be 20 'random' users)
4. Human Evaluator manually flags each selected user by looking into the available data sources (contributions graph, account links, etc)

Judgement & Sanction (end round)

1. Metadata provided by the data generators are checked against the final extrapolation results

Notes

• How the evaluator is going to handle the resulting data?
  • Right now, it is through google sheets
• Improvements
  • add github fields to the test sheet
  • The dry run will perform better by having a subset of the full contributions graph
  • Needs a better way to generate the created_on events
  • Make the generated user names less obvious

References

• Gitcoin Data Sources: [[https://hackmd.io/hVg9pm_LQN6HjJ6-uV2ZBg]]