Eleuther AI (@aieleuther) is helping make open-source AI Safety possible for everyone -- case in point: training data attribution. TDA is one of the most high-impact areas of AI research. Still, it is oddly neglected outside labs because the costs are prohibitive and implementing the methods is difficult, requiring expert mathematical understanding. As a result, although AI safety and AI academics are extremely interested in asking questions about data, most researchers never get a chance to probe a model’s internal attribution, as I did in my recent project. The tooling lived in private labs or one-off research code that never scaled. EleutherAI @aieleuther released Bergson to solve that problem. Bergson is an open library for data attribution, and this package and EleutherAI are what made our COLM paper possible. Three of Bergson's authors - Louis Jaburi, @lucia_quirke, and @blancheMinerva - are co-authors on our paper. Without @aieleuther the entire AI safety community would be unable to access this incredible resource for shaping the behaviors of AI models. Here's what it does. Every document in your training set gets run through the model and turned into a gradient, which is a record of how that document would move the weights. Bergson compresses each one and writes it to a store on disk. Then you can ask a question: here's a behavior I care about, which documents pushed the model toward it? One dot product per document gives you a signed answer. Positive means that document pushed the model toward the behavior. Negative means it pushed away. However, what Bergson does differently than the projects it was inspired by is that it is built for scale. Bergson lets you work with a gradient store on disk instead of in memory, runs under FSDP across multiple nodes, and can also handle models past 70B in our own tests. So far, our team has built indices of 5M docs from the corpora of OLMo, Marin, DCLM, SmolLM, and others. We have fully tested this software and *slaps hood of Bergson* you can fit so many methods in this baby. Gradient cosine similarity, EK-FAC, TrackStar, SOURCE, and MAGIC are all in there. Some of these are the first open implementations anyone has published! And this kind of work isn't vibe-coding. This is very precise maths that you can’t Claude up in an afternoon. Before Bergson, using MAGIC, SOURCE, or TrackStar meant building it yourself from the paper. And good luck unless you have a Math PhD! Luckily, EleutherAI has folks to do this kind of work for us. :) Bergson offers lots of methods for researchers depending on the size and budget of your task. Cheaper methods are straightforward, with simpler scoring against final checkpoints, but you also have expensive methods that model and attribute the training run itself. MAGIC is the expensive end and costs you three to five training runs worth of GPU, but it's the closest to ground truth. TrackStar scores from the final checkpoint and costs one or two training runs, but it is currently the workhorse method to reach for when your corpus has millions of documents. And switching between them is just a few config changes. Bergson is MIT-licensed, so you can run `pip install bergson` and go test it on your own favorite small language model. Paper: arxiv.org/abs/2606.11660 Code: github.com/EleutherAI/bergso… Docs: bergson.readthedocs.io