Update README.md

README.md

@@ -25,7 +25,7 @@ Note: this repo is a WIP and does not yet implement all features described below
 # Dataset Card for Proof-pile
 
 # Dataset Description
-The `proof-pile` is a 45GB pre-training dataset of mathematical text. The dataset is composed of diverse sources of both informal and formal mathematics, namely
+The `proof-pile` is a 40GB pre-training dataset of mathematical text. The dataset is composed of diverse sources of both informal and formal mathematics, namely
 - ArXiv.math (40GB)
 - Open-source math textbooks (50MB)
 - Formal mathematics libraries (500MB)
@@ -47,10 +47,15 @@ This dataset is intended to be used for pre-training language models. We envisio
 # Languages
 All informal mathematics in the `proof-pile` is written in English and LaTeX (arXiv articles in other languages are filtered out using [languagedetect](https://github.com/shuyo/language-detection/blob/wiki/ProjectHome.md)). Formal theorem proving languages represented in this dataset are Lean 3, Isabelle, Coq, HOL Light, Metamath, and Mizar.
 
- # Splits 
+ # Configurations
  The data is sorted into `"arxiv", "books", "formal", "stack-exchange", "wiki",` and `"math-dataset"` configurations. This is so that it is easy to upsample particular configurations during pre-training with the `datasets.interleave_datasets()` function. 
  
- Note that in the `"stack-exchange"`, `"wiki"`, and `"stack-exchange"` configurations, multiple documents are included in them same instance separated by the string `"<|endoftext|>"`. 
+# Evaluation
+The version of `set.mm` in this dataset has 10% of proofs replaced with `?` in order to preserve a validation and test set for Metamath provers pre-trained on the `proof-pile`. The precise split can be found here: [validation](https://github.com/zhangir-azerbayev/mm-extract/blob/main/valid_decls.json) and [test](https://github.com/zhangir-azerbayev/mm-extract/blob/main/test_decls.json). 
+
+The Lean mathlib commit used in this dataset is `6313863`. Theorems created in subsequent commits can be used for evaluating Lean theorem provers. 
+
+This dataset contains only the training set of the [MATH dataset](https://github.com/hendrycks/math). However, because this dataset contains ProofWiki, the Stacks Project, Trench's Analysis, and Stein's Number Theory, the `proof-pile` contains the validation and test sets of the [NaturalProofs dataset](https://github.com/wellecks/naturalproofs). 
 
 ## Contributions
 Authors: Zhangir Azerbayev, Edward Ayers, Bartosz Piotrowski.