ape_training / README.md
etiennebamas's picture
Upload APE training dataset
a257b8c verified
|
Raw History Blame Contribute Delete
1.43 kB
---
dataset_info:
features:
- name: uuid
dtype: string
- name: data_source
dtype: string
- name: question
dtype: string
- name: answer
dtype: string
- name: prompt_planner
dtype: string
- name: prompt_abstractor
dtype: string
splits:
- name: train
num_examples: 1799368
configs:
- config_name: default
data_files:
- split: train
path: train-*.parquet
---
# APE training data
This dataset is derived from the `train` split of
[`formalmathatepfl/sft_classic`](https://huggingface.co/datasets/formalmathatepfl/sft_classic)
at revision `21027890d0208fdc3881bb79b969e7bdf01dabbf`.
All rows with `proof_repair == true` were removed. The source contained
2,678,749 train rows; 879,381 repair rows were excluded, leaving 1,799,368
examples. The source `test` and `sft_replay_classic` splits are not included.
Each question requests an informal solution using `[Problem Restatement]`,
`[Useful Observations]`, `[Reasoning Path]`, `[Key Derivation]`, `[Proof Type]`,
`[Formal Anchors]`, and `[Final Solution]`, matching the prompt contract used by
`formalmathatepfl/reasoning_data:APE_warm_start_data`.
The `uuid`, `data_source`, and `answer` values are copied unchanged from the
source. `prompt_planner` contains the modified question followed by the exact
continuation whitespace. `prompt_abstractor` contains that question and the
verified Lean 4 proof extracted from `answer`.