Sextile Dataset: qwen3_llama3
Rank-diff sextile split for model pair:
- **Model A**: `Qwen/Qwen3-1.7B`
- **Model B**: `meta-llama/Meta-Llama-3-8B-Instruct`
## Agreement Metrics
- **MARD** (0=agree, 1=disagree): 0.2233
- **Spearman correlation**: 0.8660
## 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)… See the full description on the dataset page: https://huggingface.co/datasets/Adanato/qwen3_llama3_sextiles.
