Diagnoza: „przegrana” na CBD/DYK/GSM8K to bug harnessu, nie modelu

Observations

• base: Qwen3.5-27B (base/v3/v4) ↗
• method: pełna macierz modeli (bench_runner, n=100) + probe scoringu: TAK/NIE vs YES/NO, sum vs per-token vs kalibracja kontekstowa, accuracy vs F1 ↗
• status: clean ↗
• date: 2026-06-15 ↗

Referenced by

• GSM8K (mentions)
• GSM8K (defined-by)
• DYK (mentions)
• DYK (defined-by)
• CBD (mentions)
• CBD (defined-by)

Local graph

Provenance