An exact-deduplicated mixture of five reasoning datasets, filtered to retain the
lowest 90.00% instance perplexities under
meta-llama/Llama-3.2-1B-Instruct. The schema is id, source, question, steps, and answer.
Metric
Value
Input rows after deduplication
854,432
Retained rows
768,988
Filtered low-likelihood rows
85,444
Per-token log-likelihood cutoff
-2.49545
Maximum retained perplexity
12.1272
Source composition… See the full description on the dataset page: https://huggingface.co/datasets/cs-giung/think-sft-mini-llama3-1.2b.
