Contribution activity, refreshed automatically by GitHub Actions.
CS @ Queen's · LLM training · agentic evaluation
$ ./profile --status
focus : AI eval / data flywheels
latest : Execution Gold @ Meituan
During my internship on Meituan’s M17 Foundation Model Team – Agent, I worked on Execution Gold, a sandbox-verified data-admission protocol for agent training.
It combines 3× execution consistency, immutable artifacts, and infrastructure-failure isolation so training data has to prove it belongs in the flywheel.
Scale: 707 SWE-bench candidates screened across 15 repository families, yielding 666 evaluation slots.
📎 Metrics, with receipts
A selective-distillation pilot showed a 10.53 percentage-point difference in a small, limited-sample run. The paper treats this as directional evidence—not a universal win. I like my claims reproducible and my confidence intervals attached.
Languages: Python Java C Bash
Systems: Docker Linux Azure MCP
Meituan · Capgemini · HELLO COSMOS
Anime, board games, and badminton.