---
language:
- en
task_categories:
- text-generation
tags:
- math
- sft
- distillation-project
source_datasets:
- jasonrqh/Countdown-CoT-20k
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-10000.parquet
  - split: test
    path: data/test-500.parquet
license: mit
---

# DistillationProject: Countdown 10K training set

The exact two-phase Countdown curriculum shared by the nine Numina5K through Numina45K → Countdown10K models. The first 5,000 train rows reproduce the first Countdown phase in its actual loader order; the next 5,000 reproduce the continuation phase. The continuation used a fresh optimizer and schedule, so this dataset does not imply a single uninterrupted 10K-epoch optimizer schedule. The 500-puzzle test split is the unchanged separate Countdown benchmark and is disjoint from both train phases. Byte-identical copies of both original prepared 5K parquets are included under provenance. Training puzzle final expressions were checked with the original safe arithmetic verifier and input multiplicity rules; explanatory reasoning was not formally verified.

Exact original question and response text is preserved. No resampling, solution rewriting or truncation is performed for publication. Source revisions, row counts, order and file SHA256 values are recorded in `publication.json` and `provenance/`. Prepared tables retain their original columns. Source validity labels do not establish independently verified mathematical correctness.

Collection: https://huggingface.co/collections/zbeeb/disstillation-project-6ace05e5f40dbe715407516c