Single-Model PPL Baseline: qwen25_3b
Samples sorted by `Qwen/Qwen2.5-3B` perplexity and split into 6 equal bins.
- **Bin 0**: Lowest PPL (most familiar to the model)
- **Bin 5**: Highest PPL (most surprising to the model)
## Statistics
| Bin | Size | HQ% |
|-----|------|-----|
| 0 | 41,666 | 7.7% |
| 1 | 41,666 | 55.0% |
| 2 | 41,666 | 72.3% |
| 3 | 41,666 | 63.7% |
| 4 | 41,666 | 33.5% |
| 5 | 41,670 | 7.8% |
HQ% = percentage of samples from AceReason +… See the full description on the dataset page:
https://huggingface.co/datasets/Adanato/qwen25_3b_ppl_baseline.
