---
language: en
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
  - split: test
    path: data/test-*.parquet
---

# ELF TinyGSM — GPT-2 tokens

Processed data used by the ELF TinyGSM experiments: 11,779,329 TinyGSM training
examples and the 1,319-example GSM8K test set used for evaluation. Columns are
`index`, `input`, `target`, `condition_input_ids`, and `input_ids`. Token IDs use
the GPT-2 tokenizer. These Parquet files preserve the existing processed dataset's
values, schema, and row order; no retokenization was performed.

```python
from datasets import load_dataset

train = load_dataset("yresearch/tinygsm-gpt2", split="train")
evaluation = load_dataset("yresearch/tinygsm-gpt2", split="test")
```

Training source: [TinyGSM/TinyGSM](https://huggingface.co/datasets/TinyGSM/TinyGSM),
whose upstream dataset card declares the MIT license.

Evaluation source: [openai/gsm8k](https://huggingface.co/datasets/openai/gsm8k),
introduced by Cobbe et al. (2021), under the MIT license. Its upstream copyright
and permission notice is preserved in `LICENSE-GSM8K`.

`data_manifest.json` records the split sizes, feature types, and shard SHA-256 hashes.