ethiopian-calendar-blindspot-qwen3-4b

sudouserx

Ethiopian-Gregorian Calendar Blind-Spot Evaluation Model: Qwen/Qwen3-4BItems per stratum: 40Seed: 42Runtime: 37.2 minutes Research question Does performance on Ethiopian-calendar conversion improve when the Ethiopian-specific rules are explicitly supplied, while the same model and dates are used in closed-book and open-book conditions? The design separates three pieces of evidence: Closed-book conversion: measures the model's unaided performance. Open-book… See the full description on the dataset page: https://huggingface.co/datasets/sudouserx/ethiopian-calendar-blindspot-qwen3-4b.

ethiopian-calendar-blindspot-qwen3-4b-v4

sudouserx

Ethiopian ↔ Gregorian Calendar Blind-Spot Evaluation — v4.2_atomic_staircase Model: Qwen/Qwen3-4BItems per calendar stratum: 16Protocol: closed-book → rules-given → exact-anchor-given → exact-offset-given, thinking/non-thinking Why this version exists The previous run had very low performance on both calendar conversion and the Gregorian arithmetic control. That pattern cannot, by itself, support a claim of an Ethiopian-calendar blind spot. This release therefore… See the full description on the dataset page: https://huggingface.co/datasets/sudouserx/ethiopian-calendar-blindspot-qwen3-4b-v4.