.png&w=256&q=75)
🤓 Latest Submissions
BenchSieve
BenchSieve is a token-efficient benchmark agent designed to preserve answer quality while minimizing recorded Fireworks token usage. Instead of routing prompts only by task type, BenchSieve first detects the expected answer shape, such as a label, number, yes/no answer, JSON object, code snippet, entity list, or explanation. It then combines category classification, risk scoring, local micro-solvers, confidence checks, compact Fireworks fallback prompts, dynamic max_tokens, and output validation to produce reliable answers with fewer unnecessary tokens. The system reads benchmark tasks from /input/tasks.json, processes each prompt through a local-first routing cascade, and writes a valid /output/results.json containing only task_id and answer. Simple and low-risk tasks can be answered without remote model calls, while complex or high-risk prompts are escalated to an allowed Fireworks model using strict output budgets. BenchSieve is packaged as a Docker container for reproducible evaluation and is built for the AMD AI Developer Hackathon scoring flow: pass the accuracy gate first, then reduce token usage safely.
13 Jul 2026
