Sextile Dataset: qwen3_gemma
Rank-diff sextile split for model pair:
- **Model A**: `Qwen/Qwen3-1.7B`
- **Model B**: `google/gemma-3-1b-it`
## Agreement Metrics
- **MARD** (0=agree, 1=disagree): 0.1926
- **Spearman correlation**: 0.8944
## What are sextiles?
Data is split into 6 equal bins based on `rank_diff = rank(ppl_a) - rank(ppl_b)`:
- **Sextile 1**: Lowest diff (Model A finds these samples easiest relative to Model B)
- **Sextile 6**:… See the full description on the dataset page: https://huggingface.co/datasets/Adanato/qwen3_gemma_sextiles.
