Per-rule LoRuGEC scores

The matrix behind the paper's headline finding: per-rule F_0.5 across 35 model conditions (zero-shot, fine-tuned, continuation training). Each row is one of the 48 LoRuGEC rules.

Source: per_rule_lorugec.csv (48 × 36). Cells shaded by F0.5 value.

loading…

Reasoning vs zero-shot

Per-rule comparison of think-mode vs zero-shot for the Qwen models in the study, added in the camera-ready (R1.Q2 reviewer response).

Source: per_rule_thinking_vs_zs.csv. Cells shaded by score.

loading…