Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
| Date | Stars |
|---|---|
| 2026-07-31 | 1457 |
| 2026-08-03 | 1459 |
| 2026-08-06 | 1459 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# DataComp-LM (DCLM)
## ⚠️ Updates to centered CORE and EXTENDED calculations (9/5/2025)
**TL;DR:** Bug fixes in baseline calculations mean CORE/EXTENDED scores from before Sept 2025 are not directly comparable to new results. Rank orderings should remain consistent though.
### What happened?
As discussed in https://github.com/mlfoundations/dclm/issues/114 and addressed by https://github.com/mlfoundations/dclm/pull/115, we have recently found some bugs in the way we compute centered CORE and EXTENDED evaluation scores. At a high-level, we fixed some errors in the baseline values we assigned to some multiple choice tasks in [`eval_meta_data.csv`](./eval/eval_meta_data.csv). Since our CORE and EXTENDED averages rescale individual tasks scores relative to these baselines, our fixes WILL unfortunately cause discrepancies between older and newer results.
Generally, averaging with the corrected baselines will shift CORE/EXTENDED numbers slightly downward. However, we performed some empirical analysis that suggests rank orderings between the old and new averages should be highly consistent https://github.com/mlfoundations/dclm/pull/115.
### What this means if you are a new/current user of DCLM?
* Please use the latest version of our repo as-is but be careful when comparing to older published results.
* To compare with our results, we retain both new and old scores in [`exp_data/evals`](exp_data/evals) JSONs, storing them under the keys `{Core|Extended}_v2` and `{Core|Extended}_v1` respectively.
* To compare with evaluations that previous `dclm` users have run, you can probably assume that any results obtained prior to Sept 2025 correspond to `v1` numbers (but please double check with the authors if unsure).
* If you wish to compute `v1` averages yourself, you can run the script [`eval/aggregated_metrics.py`](eval/aggregated_metrics.py) on any JSON that your own evaluations have generated. This will add in entries for `Core_v1` and `Extended_v1`.
```
python eval/aggregated_metrics.py \
--eval_results <PATH_TO_YOUR_EVAL_JSON> \
--version v1
```
### What this means if you have previously used DCLM?
* Your rank orderings and conclusions are unlikely to change!
* However, you can also update your scores via the script [`eval/aggregated_metrics.py`](eval/aggregated_metrics.py). This will update all averages to be `v2` while migrating existing `Core` and `Extended` keys to `Core_v1` and `Extended_v1` respectively.
* For scientific clarity, we'd really appreciate it if you could do so (or at least mark whether your reported scores are `v1` or `v2`).
```
python eval/aggregated_metrics.py \
--eval_results <PATH_TO_YOUR_EVAL_JSON> \
--version v2 # (default)
```
To all, we sincerely apologize for any inconveniences that these fixes may cause.
## Table of Contents
- [Introduction](#introduction)
- [Leaderboard](#leaderboard)
- [Getting Started](#getting-started)
- [Workflow Overview (and Exp Data)](#workflow-overview-and-exp-data)
- [Selecting Raw Sources](#1-selecting-raw-sources)
- [Processing the Data](#2-processing-the-data)
- [Ray-based Filtering and Cleaning](#ray-based-filtering-and-cleaning)
- [Deduplication](#deduplication)
- [Tokenization and Shuffling](#3-tokenization-and-shuffling)
- [Model Training](#4-model-training)
- [Model Evaluation](#5-model-evaluation)
- [Submission](#submission)
- [Downloading Artifacts](#downloading-artifacts)
- [Datasets](#datasets)
- [Pretrained Models](#pretrained-models)
- [Example Commands to Train 1B model with DCLM baseline](#example-commands-to-train-1b-model-with-dclm-baseline)
- [How to Contribute](#how-to-contribute)
- [How to Cite Us](#how-to-cite-us)
- [License](#license)
## Introduction
[DataComp-LM (DCLM)](https://datacomp.ai/dclm/) is a comprehensive framework designed for building and training large language models (LLMs) with diverse datasets. It offers a standardized corpus of over 300T unfiltered tokens from CommonCrawl, effecExcerpt of 40,539 characters
Read on GitHub
17
16
14
11
10
4
3
2
Ryan Marten
1
Achal Dave · Toyota Research Institute
1
Ludwig Schmidt · Stanford University, Anthropic · United States
1
David Wadden · University of Washington · United States
1
Niklas
1
1
1
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:8237ed9fc8337fe0, llm:Repository name and description: 'mlfoundations/dclm' — "DataComp for Language Models" indicating a dataset/benchmark for language models.
matched fp:8237ed9fc8337fe0, llm:Repository name and description: 'mlfoundations/dclm' — "DataComp for Language Models" indicating a dataset/benchmark for language models.