What it does
On its RACE-H benchmark, writing each passage once for four questions cut token use 2.5-fold and raised throughput 2.3-fold.
Maker-reported (not independently measured by JevMade): RACE-H, 250 passages x 4 questions (n = 1000). One passage, four multiple-choice questions · Packing writes the passage once instead of four times: 2.5x fewer tokens, 2.3x the throughput of batching,
- Primitives
- choice
- Platform
- Python
- Added
- Project created
- GitHub stars
- 3 · snapshot 2026-09-18T22:58:45Z
Source checked 2026-09-19 — opened the GitHub repository directly.