SlavaYus/tinysms_train_prompts_12m

tinysms train prompts 12m

Every training row of TinyGSM/TinyGSM as a math question and its reference program, preprocessed exactly as the model is trained on them: the docstring is stripped and the entry point is renamed to solve. Nothing else is stored -- no answers, no token counts, no rendered prompt, no ids.

• train: 11,838,003 rows in 60 parquet file(s)
• test: the 1,919 questions of the project's test split -- questions only, no reference programs and no gold values, so the split is safe to publish
• the last 4,000 rows of shard 0 -- the validation and test rows of src/data.py -- are not in train
• 4,106 further rows whose question text is one of the val/test questions were dropped as well (DROP_EVAL_COPIES = True): TinyGSM ships the same question with several reference programs, so those rows are copies of the evaluation questions
• the reference programs are not executed, unlike the smaller tinysms_train_qa and tinysms_train_qa_1m sets: no answers are shipped, so a reference that fails to return a number is not filtered out
• built shard by shard: every parquet file is uploaded to the Hub and deleted before the next one is written, so the whole set never exists on disk at once

Loading

from datasets import load_dataset
ds = load_dataset("SlavaYus/tinysms_train_prompts_12m")           # DatasetDict(train=..., test=...)
print(ds["train"][0], len(ds["test"]))

Schema

train:

question: string

test:

question: string