Datasets Matter: The Battle Between Open and Closed Generative AI is Not Only About Models Anymore

Two major open source datasets were released this week.

Created Using DALL-E

Next Week in The Sequence:

  • Edge 403: Our series about autonomous agents continues covering memory-based planning methods. The research behind the TravelPlanner benchmark for planning in LLMs and the impressive MemGPT framework for autonomous agents.

  • The Sequence Chat: A super cool interview with one of the engineers behind Azure OpenAI Service and Microsoft CoPilot.

  • Edge 404: We dive into Meta AI’s amazing research for predicting multiple tokens at the same time in LLMs.

You can subscribe to The Sequence below:

TheSequence is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

📝 Editorial: Datasets Matter: The Battle Between Open and Closed Generative AI is Not Only About Models Anymore

The battle between open and closed generative AI has been at the center of industry developments. From the very beginning, the focus has been on open vs. closed models, such as Mistral and Llama vs. GPT-4 and Claude. Less attention has been paid to other foundational aspects of the model lifecycle, such as the datasets used for training and fine-tuning. In fact, one of the limitations of the so-called open weight models is that they don’t disclose the training datasets and pipeline. What if we had high-quality open source datasets that rival those used to pretrain massive foundation models?

Open source datasets are one of the key aspects to unlocking innovation in generative AI. The costs required to build multi-trillion token datasets are completely prohibitive to most organizations. Leading AI labs, such as the Allen AI Institute, have been at the forefront of this idea, regularly open sourcing high-quality datasets such as the ones used for the Olmo model. Now it seems that they are getting some help.

This week, we saw two major efforts related to open source generative AI datasets. Hugging Face open-sourced FineWeb, a 44TB dataset of 15 trillion tokens derived from 96 CommonCrawl snapshots. Hugging Face also released FineWeb-Edu, a subset of FineWeb focused on educational value. But Hugging Face was not the only company actively releasing open source datasets. Complementing the FineWeb release, AI startup Zyphra released Zyda, a 1.3 trillion token dataset for language modeling. The construction of Zyda seems to have focused on a very meticulous filtering and deduplication process and shows remarkable performance compared to other datasets such as Dolma or RedefinedWeb.

High-quality open source datasets are paramount to enabling innovation in open generative models. Researchers using these datasets can now focus on pretraining pipelines and optimizations, while teams using those models for fine-tuning or inference can have a clearer way to explain outputs based on the composition of the dataset. The battle between open and closed generative AI is not just about models anymore.

🔎 ML Research

Extracting Concepts from GPT-4

OpenAI published a paper proposing an interpretability technique to understanding neural activity within LLMs. Specifically, the method uses k-sparse autoencoders to control sparsity which leads to more interpretable models —> Read more.

Transformer are SSMs

Researchers from Princeton University and Carnegie Mellon University published a paper outlining theoretical connections between transformers and SSMs. The paper also proposes a framework called state space duality and a new architecture called Mamba-2 which improves the performance over its predecessors by 2-8x —> Read more.

Believe or Not Believe LLMs

Google DeepMind published a paper proposing a technique to quantify uncertainty in LLM responses. The paper explores different sources of uncertainty such as lack of knowledge and randomness in order to quantify the reliability of an LLM output —> Read more.

CodecLM

Google Research published a paper introducing CodecLM, a framework for using synthetic data for LLM alignment in downstream tasks. CodecLM leverages LLMs like Gemini to encode seed intrstructions into the metadata and then decodes it into synthetic intstructions —> Read more.

TinyAgent

Researchers from UC Berkeley published a detailed blog post about TinyAgent, a function calling tuning method for small language models. TinyAgent aims to enable function calling LLMs that can run on mobile or IoT devices —> Read more.

Parrot

Researchers from Shanghai Jiao Tong University and Microsoft Research published a paper introducing Parrot, a framework for correlating multiple LLM requests. Parrot uses the concept of a Semantic Variable to annotate input/output variables in LLMs to enable the creation of a data pipeline with LLMs —> Read more.

🤖 Cool AI Tech Releases

FineWeb

HuggingFace open sourced FineWeb, a 15 trillion token dataset for LLM training —> Read more.

Stable Audion Open

Stability AI open source Stable Audio Open, its new generative audio model —> Read more.

Mistral Fine-Tune

Mistral open sourced mistral-finetune SDK and services for fine-tuning models programmatically —> Read more.

Zyda

Zyphra Technologies open sourced Zyda, a 1.3 trillion token dataset that powers the version of its Zamba models —> Read more.

🛠 Real World AI

Salesforce discusses their use of Amazon SageMaker in their Einstein platform —> Read more.

📡AI Radar

TheSequence is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

By Julio MarchiPosted on