Andriy Mulyar on GPT4All, Data Quality and Open Local Models

Matthew Berman introduces Andriy Mulyar, co-founder of Nomic AI, the company behind the Atlas data exploration tool and GPT4All. Mulyar argues that high-quality, clean data is often more important to a good machine-learning model than a new architecture.

Mulyar was born in Lviv, Ukraine, and moved with his family to Richmond, Virginia, at age four after they won the US green card lottery. He learned Python in middle school, built a Minecraft server, and attended a high school focused on computing and mathematics.

In college he joined an NLP lab, published research and worked on language models for medical data before beginning a PhD in interpretable machine learning. He left the program to found Nomic, believing he could make a larger difference by building openly.

His early research involved named entity recognition: extracting structured information from text. An internship with Mark Dredze at Johns Hopkins introduced him to language models and helped set his direction in AI.

At a medical startup, Mulyar helped train a model to summarize data for radiologists. Hallucination was a serious obstacle in medicine, where invented details can harm patients. He says intensive data quality and cleaning work helped the team address the problem.

That experience convinced him that curated data is central to model quality. Later, as interest in language models grew, he and a former colleague decided to build tools that would make it easier to inspect and curate training data.

Berman asks about spurious correlations. Mulyar describes them as shortcuts a model finds in its training data that should not determine the intended outcome. A credit model, for example, might use ZIP code to accept or reject applicants, potentially marginalizing whole groups.

Language models can inherit similar correlations from text. To assess model quality, Mulyar describes evaluating on unseen data, using quantitative measures where possible and human judgments with statistical comparisons for qualitative behavior. He says automatic evaluation is especially difficult when the output is open-ended text.

Atlas was built to help people examine large text collections before an expensive training run. Instead of manually reading a small sample from a million documents, users can inspect a pre-organized map of the collection and ask domain experts what should be removed or changed.

Mulyar explains that Atlas converts each uploaded document into an embedding vector using machine-learning models, then projects the high-dimensional vectors into two dimensions. Documents the model considers similar appear near one another, offering a human-readable view of the data.

Developers and data scientists use Atlas to understand large collections of text or images, while librarians can explore digitized archives. Users can filter and manipulate the data and share a link to an interactive map. At the time, Mulyar says a free tier included 50,000 data points, with larger allocations for researchers.

Mulyar describes GPT4All as a language model that runs locally without a GPU or a particularly fast computer. It was less capable than OpenAI’s cloud models at the time, he says, but could perform many similar tasks without sending data to a third party.

Nomic had been building Atlas before the public surge of interest in language models in late 2022. The team wanted to demonstrate that its tool could curate high-quality data, while making a model accessible to people who lacked expensive GPUs or could not access cloud APIs in their countries.

Meta’s LLaMA model provided a foundation for fine-tuning, while Georgi Gerganov’s CPU work showed that LLaMA could run quickly without a GPU. Seeing Alpaca’s success, Nomic used Atlas to curate data for an assistant model and released GPT4All.

Mulyar credits the project’s simple one-line installation and clear instructions with removing friction. The release demonstrated that language models could run on a personal computer, not only behind a cloud API.

Berman says he was surprised by the reception to his own GPT4All video and asks whether Nomic expected such adoption. Mulyar says it did not: the original goal was to demonstrate the value of clean data, but the response revealed substantial demand for accessible local models.

Mulyar says organizations were concerned about sending private data to third parties, while others needed offline access. Berman adds that local access lets people experiment with, extend and fine-tune models instead of relying only on a closed API.

Nomic released the training data as well as other components. Mulyar argues that public scrutiny can help the technology develop in a safer direction, though he notes that many model developers do not release their data, including because of copyright concerns.

The original GPT4All was fine-tuned from LLaMA using responses collected from an OpenAI assistant API. Mulyar says the original model was for research and could not be used commercially under the base model’s license; he also discusses OpenAI’s competing-product terms.

Demand from people who wanted to build businesses on local models pushed the team to seek a more permissive foundation. They chose GPT-J, trained on the Pile, as the base for GPT4All-J. Mulyar notes that the Pile contains internet material and that training-data copyright raised unresolved legal questions.

GPT-J was released by EleutherAI under Apache 2.0, but Mulyar says it was weaker than LLaMA. To improve the assistant, Nomic increased and curated its fine-tuning data, using Atlas to debug problems such as memorization rather than generalization.

Mulyar credits a Discord community of roughly 8,000 to 9,000 people with helping release the new model. Contributors suggested data and donated development time; one adapted low-level CPU code so that GPT-J could run quickly without a GPU.

Asked about GPU support, Mulyar explains that quantization shrinks the original model weights for CPU use. Nomic also released the larger raw weights, which technically capable users could run on a GPU; GPU support in the chat launcher was on the short-term roadmap.

Berman asks why installers generated unverified-developer warnings. Mulyar says the team was waiting for Apple and Microsoft signing certificates and had chosen to release early and gather feedback rather than wait for those approvals.

Fine-tuning raises several different questions, Mulyar says: whether a model can be trained further, whether the result improves, and whether it can learn from private data while preserving its earlier behavior. GPT4All-J training instructions were in the project’s GitHub README.

At that time he says training still required substantial GPU resources and memory for the raw weights. Berman suggests that many people who ask for fine-tuning may instead want to ask questions about their own documents.

Mulyar agrees that retrieval-augmented generation can be a better fit for that task. Relevant information can be provided in the prompt before the question, and with larger document collections a retrieval system can select context. He points to LangChain tutorials and its GPT4All integration.

For actual fine-tuning, Mulyar describes renting an eight-GPU DGX A100 and preparing properly formatted examples. He estimates thousands of conversations may be needed, such as a PDF excerpt, a question about it and a written answer. He discloses that Paperspace helped sponsor GPT4All’s training.

On multilingual support, Mulyar says the base model must learn from diverse languages during pretraining, and instruction fine-tuning data must also include questions and answers in those languages. More language coverage was on the project’s roadmap at the time.

He explains that a token is not the same as a word and that the model’s context limit constrains what it can consider. The version under discussion had a 2,048-token limit covering the prompt and response together, including prior conversation; longer contexts were planned.

Berman asks how model bias arises. Mulyar says models inherit patterns and biases in their training data, just as a credit model may pick up undesirable correlations. As an example, he says language models may associate nurses more often with women and plumbers more often with men.

Because the original GPT4All training examples came from another model’s API, Mulyar says its biases can carry into the derived model. Publishing the dataset lets the community inspect and change it, but he acknowledges that people could also intentionally create harmful biases and deploy them in automated systems.

Mulyar argues that openness enables scrutiny and defenses against harmful uses, while concentrating control in one organization creates a different risk. Berman asks whether biases arise from internet data or decisions by model developers.

Mulyar distinguishes patterns inherited from pretraining data from choices made when curating assistant training. Refusals to help with illegal acts or uncertain medical advice are examples of deliberately shaped behavior. He notes that such restrictions can also block benign uses, such as fictional medical advice in a game.

He cautions against a blanket claim that local models have more or less bias than cloud models: the answer depends on the model and its training data. The original GPT4All dataset was imperfect, he says, and community members used its Atlas map to identify examples they considered problematic.

Berman asks whether filtering and bias are separate. Mulyar considers filtering a way of changing a model’s biases: removing training examples with a particular refusal, for instance, could make a model less likely to refuse similar requests. Every model, in his view, has some bias.

He says users found the newly released model too restrictive, so the team had started training a less restrictive version. He presents the quick change as an example of open data enabling interventions, while acknowledging that openness also creates opportunities for misuse.

Mulyar gives a hypothetical “bring your own bias” example: a company could pay to replace references to a rival drink with its own brand in training data. He argues that access to datasets could help researchers build ways to detect or mitigate such manipulation.

Turning to centralization, Mulyar credits major model providers with making useful AI assistants and enabling applications. He worries, however, that the field has become less open as commercial pressures grow.

He says OpenAI previously published more information about models, but its GPT-4 report omitted training details. That decision helped motivate him to build GPT4All: he wanted people to have alternatives and access to the technology without needing an internet connection.

Berman raises the argument that widely available models are easier to misuse. Mulyar accepts that risk but also worries about concentrated corporate power and incentives; he favors putting models in more people’s hands. Berman suggests centralized and local models can coexist and keep each other accountable.

Mulyar says the local models available in 2023 were generally less capable and harder to use than leading cloud systems, but offered offline access, privacy and reduced dependence on a provider’s future prices. Companies concerned about employees sending sensitive data to cloud services could use local models instead.

Asked how local models could catch up, Mulyar emphasizes high-quality assistant interaction data, including edited examples that reinforce good responses. He says GPT4All users could opt in to contribute conversations to a public, downloadable data lake; sharing was not enabled by default.

Berman asks whether contributors could be identified. Mulyar describes options to associate data with a contributor for credit or anonymize it, and says users in Europe could delete submitted data. These are his descriptions of the project’s practices at the time.

Mulyar says another major ingredient is expensive pretraining of a very large model, an area where well-funded organizations have an advantage. He credits OpenAI’s contributions and argues that a strong openly released foundation model would help local systems approach the leading cloud models.

For unusual deployments, Mulyar describes a model needing roughly 4 GB of RAM running on a small CPU connected to a TI-84 calculator or a Nintendo DSi. The calculator acts as an interface; it does not itself hold and run the model. Nomic also offered community bounties for creative edge-device uses.

He suggests an offline phone could run a model and answer questions away from an internet connection. Berman marvels at access to information through everyday devices, while Mulyar credits the many preceding contributions that made the recent progress possible.

Berman asks about the 2023 open letter calling for a six-month pause in training systems more powerful than GPT-4. Mulyar says he only skimmed it and lacks a strong view on the letter itself, though he is concerned when people with financial incentives attempt to block progress.

On AI’s longer-term role, Mulyar predicts that calls to language models will become a basic part of interacting with computers and that community work will improve models on smaller devices. He expects open models to let people perform useful tasks without always sending data to a third party.

Berman thanks Mulyar and points viewers to Atlas and GPT4All. Mulyar says the GitHub project was the best place to find the nascent ecosystem and follow its development.