The reviewed record of science sign in
Pith

arxiv: 2606.11660 · v1 · pith:X3FC62YK · submitted 2026-06-10 · cs.LG

Bergson: An Open Source Library for Data Attribution

Lucia Quirke , Louis Jaburi , David Johnston , William Z. Li , Gon\c{c}alo Paulo , Guillaume Martres , Girish Gupta , Stella Biderman
show 1 more author
Nora Belrose
This is my paper

Reviewed by Pith2026-06-27 10:46 UTCgrok-4.3pith:X3FC62YKopen to challenge →

classification cs.LG
keywords data attributionmodel interpretabilitylarge language modelsopen source libraryMAGICSOURCETrackStargradient stores
0
0 comments X

The pith

Bergson supplies the first open-source implementations of MAGIC, SOURCE, and TrackStar that scale to very large language models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Data attribution traces how individual training examples shape a model's outputs, which can help identify sources of unwanted behavior or guide better dataset choices. The paper presents Bergson as a library built to run these techniques on the scale of current large language models and their pre-training data. It adds native support for storing gradients on disk and spreading work across multiple machines, plus everyday tools that reduce setup time. The release also ships the first public versions of three established methods named MAGIC, SOURCE, and TrackStar. A reader would care because prior work in this area stayed limited to small models or stayed behind closed implementations.

Core claim

Bergson is an open source library that aims to enable faster progress in the field by providing a host of techniques that scale to very large language models and pre-training datasets. The library natively supports on-disk gradient stores and multi-node distributed training, and provides quality of life tools for researchers. Finally, it introduces the first open-source implementations of three leading data attribution methods: MAGIC, SOURCE, and TrackStar.

What carries the argument

The Bergson library, which performs data attribution through on-disk gradient stores and multi-node distributed training to reach the scale of large language models and pre-training datasets.

If this is right

  • Researchers gain immediate access to open code for MAGIC, SOURCE, and TrackStar without writing their own versions.
  • Data attribution work on full pre-training datasets becomes practical through the library's built-in distributed training and disk storage.
  • Quality of life tools in the library lower the engineering cost of running attribution experiments.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Routine use of the library could turn data attribution into a standard check applied during model development.
  • The shared code base may let groups compare different attribution methods on identical large-scale setups.
  • Community additions of further methods to the library could expand the set of available techniques over time.

Load-bearing premise

The library's versions of MAGIC, SOURCE, and TrackStar reproduce the original methods and retain their performance when applied to very large models and datasets.

What would settle it

Running Bergson's implementations and the original method code side by side on one large model and dataset, then checking whether the attribution scores match within expected numerical tolerance.

Figures

Figures reproduced from arXiv: 2606.11660 by David Johnston, Girish Gupta, Gon\c{c}alo Paulo, Guillaume Martres, Louis Jaburi, Lucia Quirke, Nora Belrose, Stella Biderman, William Z. Li.

Figure 1
Figure 1. Figure 1: Bergson’s composable pipeline. We provide a fully functional, differentiable, scalable, and deterministic trainer compatible with all supported attribution methods, but general-purpose trainers may also be used to attribute with influence functions (main pipeline). Hessian options include autocorrelation, K-FAC, EK-FAC, TK-FAC/Shampoo, and identity. Normalization options include Adam or Adafactor second mo… view at source ↗
Figure 2
Figure 2. Figure 2: TrackStar pipeline declared as a YAML file. bergson <file-name> executes the file, performing a composition of empirical Fisher information matrix fitting and mixing, query￾index building, and query scoring. & Engstrom, 2025). These scores may be of interest as a proxy for ground truth leave-one-out influences during method development, as an alternative to leave-k-out re￾training iterations. To support th… view at source ↗
Figure 3
Figure 3. Figure 3: Token level attribution for biosecurity capabilities. The tokens highlighted in red are predicted to improve model perfor￾mance on the robust subset of the WMDP Bio evaluation, while the tokens highlighted in blue suppress performance. provides a pre-configured example demonstrating the neces￾sary setup. 3. Methods Guide The problem of data attribution could in principle be solved exactly by computing the … view at source ↗
Figure 4
Figure 4. Figure 4: Attributing emergent misalignment behavior using a GRPO loss. An LLM judge assigns scores to each model genera￾tion in a batch. Using Bergson’s GRPO capabilities, the training data that most contributed to this behavior can be found. 2025). We use data attribution to produce token level attri￾butions for biosecurity proxy knowledge, then validate the attributions with re-training experiments and visualize … view at source ↗
Figure 5
Figure 5. Figure 5: Tracing factual knowledge across writing styles. The training data is split with one portion written in Shakespearean style and another in legalese. The query set contains queries in the legalese style, while the answers are in the Shakespearean portion of the training data. To understand whether data influences can be decomposed, we created synthetic biographical facts (e.g., "Alice works at Fermilab") fo… view at source ↗
Figure 6
Figure 6. Figure 6: End-to-end attribution time vs. number of train tokens. Each run uses the same single GPU, base model, and sample sequence length. The time taken to train the base model for an equivalent number of tokens is provided as a point of comparison [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Gradient collection latency by number of GPUs em￾ployed for three models from the Pythia suite. Gradients for the smaller models are collected using replicated data parallelism, while the parallelism strategy for the larger model shifts from sharded data parallelism when collecting gradients on a single￾node, to a mixture of sharded and replicated data parallelism using multiple nodes. 6. Related Work The … view at source ↗
read the original abstract

Data attribution is a promising field in interpretability that aims to explain model behavior through the influence of its training data, with applications including debugging undesirable model behavior and training dataset curation. However, significant engineering effort is required to perform it at scale, and many cutting edge techniques lack open-source tooling and support. Bergson is an open source library that aims to enable faster progress in the field by providing a host of techniques that scale to very large language models and pre-training datasets. The library natively supports on-disk gradient stores and multi-node distributed training, and provides quality of life tools for researchers. Finally, we introduce the first open-source implementations of three leading data attribution methods: MAGIC, SOURCE, and TrackStar. The library is available at https://github.com/EleutherAI/bergson .

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The manuscript introduces Bergson, an open-source library for data attribution in machine learning. It aims to reduce engineering effort for applying attribution methods at scale on large language models and pre-training datasets by providing native support for on-disk gradient stores and multi-node distributed training, along with quality-of-life tools. The paper claims to deliver the first open-source implementations of the MAGIC, SOURCE, and TrackStar methods.

Significance. If the implementations are faithful to the originals and achieve the claimed scalability, the library would lower barriers for data attribution research, enabling wider use of these techniques for model debugging and dataset curation. The open-source release itself is a concrete contribution that could accelerate progress in interpretability.

major comments (1)
  1. [Abstract] Abstract: The central claims that Bergson 'provides a host of techniques that scale to very large language models and pre-training datasets' and supplies 'the first open-source implementations of three leading data attribution methods: MAGIC, SOURCE, and TrackStar' are asserted without any benchmarks, timing results, memory measurements, scaling curves, or side-by-side verification of faithfulness to the source papers. The manuscript provides no empirical section demonstrating these properties and instead points to the external repository.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their thoughtful review and constructive feedback on our manuscript describing the Bergson library. We address the major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claims that Bergson 'provides a host of techniques that scale to very large language models and pre-training datasets' and supplies 'the first open-source implementations of three leading data attribution methods: MAGIC, SOURCE, and TrackStar' are asserted without any benchmarks, timing results, memory measurements, scaling curves, or side-by-side verification of faithfulness to the source papers. The manuscript provides no empirical section demonstrating these properties and instead points to the external repository.

    Authors: We acknowledge the referee's observation that the abstract makes claims about scalability and the novelty of the implementations without accompanying empirical results in the manuscript itself. However, Bergson is presented as a software library paper whose primary contribution is the release of production-ready, scalable tooling rather than new methodological advances or performance benchmarks. The scalability claims are grounded in the library's explicit design choices (on-disk gradient stores and multi-node distributed training support), which directly address the engineering barriers described in the introduction. The implementations of MAGIC, SOURCE, and TrackStar are released as open source precisely so that faithfulness can be verified by direct code inspection against the source papers; we believe this is the appropriate form of validation for a library release. Adding an empirical section would shift the paper's focus away from its intended purpose as a tool description. We therefore do not plan to revise the manuscript to include benchmarks or timing results. revision: no

Circularity Check

0 steps flagged

No circularity: software library announcement contains no derivations or predictions

full rationale

The manuscript is a library release note. It contains no equations, no claimed first-principles derivations, no fitted parameters presented as predictions, and no self-citation chains used to justify mathematical results. The central statements concern the existence and features of the Bergson library (on-disk stores, distributed training, open-source implementations of MAGIC/SOURCE/TrackStar). These are factual assertions about code, not reductions of one quantity to another by construction. No load-bearing step reduces to its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is a software library announcement rather than a theoretical paper; no free parameters, axioms, or invented entities are introduced.

pith-pipeline@v0.9.1-grok · 5687 in / 1044 out tokens · 28350 ms · 2026-06-27T10:46:59.476927+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 4 canonical work pages

  1. [1]

    Advances in Neural Information Processing Systems , editor =

    Deng, Junwei and Li, Ting-Wei and Zhang, Shiyuan and Liu, Shixuan and Pan, Yijun and Huang, Hao and Wang, Xinhe and Hu, Pingbang and Zhang, Xingjian and Ma, Jiaqi , title =. Advances in Neural Information Processing Systems , editor =

  2. [2]

    2024 , eprint=

    MATES: Model-Aware Data Selection for Efficient Pretraining with Data Influence Models , author=. 2024 , eprint=

    2024

  3. [3]

    2025 , eprint=

    Procedural Knowledge in Pretraining Drives Reasoning in Large Language Models , author=. 2025 , eprint=

    2025

  4. [4]

    2020 , eprint=

    Explaining Black Box Predictions and Unveiling Data Artifacts through Influence Functions , author=. 2020 , eprint=

    2020

  5. [5]

    2020 , eprint=

    Understanding Black-box Predictions via Influence Functions , author=. 2020 , eprint=

    2020

  6. [6]

    2025 , eprint=

    Understanding R1-Zero-Like Training: A Critical Perspective , author=. 2025 , eprint=

    2025

  7. [7]

    2026 , month = apr, day =

    Grosse, Roger , title =. 2026 , month = apr, day =

    2026

  8. [8]

    2025 , eprint=

    Optimizing ML Training with Metagradient Descent , author=. 2025 , eprint=

    2025

  9. [9]

    2022 , eprint=

    If Influence Functions are the Answer, Then What is the Question? , author=. 2022 , eprint=

    2022

  10. [10]

    International Conference on Machine Learning , pages=

    Modeldiff: A framework for comparing learning algorithms , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  11. [11]

    Utility: Do We Have to Sacrifice One for the Other in Data Importance Quantification? , author=

    Scalability vs. Utility: Do We Have to Sacrifice One for the Other in Data Importance Quantification? , author=. 2021 , eprint=

    2021

  12. [12]

    2024 , eprint=

    Scalable Influence and Fact Tracing for Large Language Model Pretraining , author=. 2024 , eprint=

    2024

  13. [13]

    2022 , eprint=

    Towards Tracing Factual Knowledge in Language Models Back to the Training Data , author=. 2022 , eprint=

    2022

  14. [14]

    An Empirical Study of Memorization in NLP

    Zheng, Xiaosen and Jiang, Jing. An Empirical Study of Memorization in NLP. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022. doi:10.18653/v1/2022.acl-long.434

    work page doi:10.18653/v1/2022.acl-long.434 2022

  15. [15]

    2023 , eprint=

    Studying Large Language Model Generalization with Influence Functions , author=. 2023 , eprint=

    2023

  16. [16]

    2025 , eprint=

    Tulu 3: Pushing Frontiers in Open Language Model Post-Training , author=. 2025 , eprint=

    2025

  17. [17]

    2025 , eprint=

    A Snapshot of Influence: A Local Data Attribution Framework for Online Reinforcement Learning , author=. 2025 , eprint=

    2025

  18. [18]

    2024 , eprint=

    LESS: Selecting Influential Data for Targeted Instruction Tuning , author=. 2024 , eprint=

    2024

  19. [19]

    TRAK: attributing model behavior at scale , year =

    Park, Sung Min and Georgiev, Kristian and Ilyas, Andrew and Leclerc, Guillaume and M. TRAK: attributing model behavior at scale , year =. Proceedings of the 40th International Conference on Machine Learning , articleno =

  20. [20]

    2018 , eprint=

    Representer Point Selection for Explaining Deep Neural Networks , author=. 2018 , eprint=

    2018

  21. [21]

    2021 , eprint=

    Input Similarity from the Neural Network Perspective , author=. 2021 , eprint=

    2021

  22. [22]

    2024 , eprint=

    Sometimes I am a Tree: Data Drives Unstable Hierarchical Generalization , author=. 2024 , eprint=

    2024

  23. [23]

    2023 , eprint=

    Progress measures for grokking via mechanistic interpretability , author=. 2023 , eprint=

    2023

  24. [24]

    2025 , eprint=

    The Faiss library , author=. 2025 , eprint=

    2025

  25. [25]

    arXiv preprint arXiv:2503.13751 , year=

    Optimizing ml training with metagradient descent , author=. arXiv preprint arXiv:2503.13751 , year=

    work page arXiv

  26. [26]

    2025 , eprint=

    MAGIC: Near-Optimal Data Attribution for Deep Learning , author=. 2025 , eprint=

    2025

  27. [27]

    2023 , eprint=

    Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling , author=. 2023 , eprint=

    2023

  28. [28]

    Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations , pages=

    Transformers: State-of-the-art natural language processing , author=. Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations , pages=

    2020

  29. [29]

    Theoretical Economics Letters , volume=

    A value for n-person games , author=. Theoretical Economics Letters , volume=. 1953 , publisher=

    1953

  30. [30]

    2020 , eprint=

    Estimating Training Data Influence by Tracing Gradient Descent , author=. 2020 , eprint=

    2020

  31. [31]

    2025 , eprint=

    Distributional Training Data Attribution: What do Influence Functions Sample? , author=. 2025 , eprint=

    2025

  32. [32]

    2024 , eprint=

    Training Data Attribution via Approximate Unrolled Differentiation , author=. 2024 , eprint=

    2024

  33. [33]

    Advances in Neural Information Processing Systems , volume=

    Manipulating sgd with data ordering attacks , author=. Advances in Neural Information Processing Systems , volume=

  34. [34]

    Mechanistic Interpretability Workshop at NeurIPS 2025 , year=

    Mitigating Emergent Misalignment with Data Attribution , author=. Mechanistic Interpretability Workshop at NeurIPS 2025 , year=

    2025

  35. [35]

    IEEE Transactions on Big Data , year=

    The faiss library , author=. IEEE Transactions on Big Data , year=

  36. [36]

    Embedding-based Retrieval in Facebook Search , url=

    Huang, Jui-Ting and Sharma, Ashish and Sun, Shuying and Xia, Li and Zhang, David and Pronin, Philip and Padmanabhan, Janani and Ottaviano, Giuseppe and Yang, Linjun , year=. Embedding-based Retrieval in Facebook Search , url=. doi:10.1145/3394486.3403305 , booktitle=

    work page doi:10.1145/3394486.3403305

  37. [37]

    2023 , eprint=

    Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena , author=. 2023 , eprint=

    2023

  38. [38]

    2022 , eprint=

    Emergent Abilities of Large Language Models , author=. 2022 , eprint=

    2022

  39. [39]

    2024 , eprint=

    SimpleFSDP: Simpler Fully Sharded Data Parallel with torch.compile , author=. 2024 , eprint=

    2024

  40. [40]

    2018 , eprint=

    Measuring the Intrinsic Dimension of Objective Landscapes , author=. 2018 , eprint=

    2018

  41. [41]

    Advanced Neural Computers , editor =

    Hecht-Nielsen, Robert , title =. Advanced Neural Computers , editor =

  42. [42]

    2026 , eprint=

    Olmo 3 , author=. 2026 , eprint=

    2026

  43. [43]

    2023 , eprint=

    LLaMA: Open and Efficient Foundation Language Models , author=. 2023 , eprint=

    2023

  44. [44]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  45. [45]

    2026 , eprint=

    Deep Ignorance: Filtering Pretraining Data Builds Tamper-Resistant Safeguards into Open-Weight LLMs , author=. 2026 , eprint=

    2026

  46. [46]

    2024 , eprint=

    The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning , author=. 2024 , eprint=

    2024

  47. [47]

    2024 , eprint=

    DataInf: Efficiently Estimating Data Influence in LoRA-tuned LLMs and Diffusion Models , author=. 2024 , eprint=

    2024

  48. [48]

    International Conference on Machine Learning , year=

    Deep learning via Hessian-free optimization , author=. International Conference on Machine Learning , year=

  49. [49]

    2021 , eprint=

    Scaling Up Influence Functions , author=. 2021 , eprint=

    2021

  50. [50]

    2017 , eprint=

    Second-Order Stochastic Optimization for Machine Learning in Linear Time , author=. 2017 , eprint=

    2017

  51. [51]

    2021 , eprint=

    LoRA: Low-Rank Adaptation of Large Language Models , author=. 2021 , eprint=

    2021

  52. [52]

    doi:10.5281/zenodo.5371629 , version =

    A framework for few-shot language model evaluation. doi:10.5281/zenodo.5371629 , version =

    work page doi:10.5281/zenodo.5371629

  53. [53]

    2024 , eprint=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. 2024 , eprint=

    2024

  54. [54]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Causal estimation of memorisation profiles , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  55. [55]

    Sourab Mangrulkar and Sylvain Gugger and Lysandre Debut and Younes Belkada and Sayak Paul and Benjamin Bossan and Marian Tietz , howpublished =

  56. [56]

    2025 , eprint=

    Better Training Data Attribution via Better Inverse Hessian-Vector Products , author=. 2025 , eprint=

    2025

  57. [57]

    2020 , eprint=

    Language Models are Few-Shot Learners , author=. 2020 , eprint=

    2020

  58. [58]

    2026 , eprint=

    Shaping capabilities with token-level data filtering , author=. 2026 , eprint=

    2026

  59. [59]

    2025 , eprint=

    Not Every Token Needs Forgetting: Selective Unlearning to Limit Change in Utility in Large Language Model Unlearning , author=. 2025 , eprint=

    2025

  60. [60]

    2024 , eprint=

    WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs , author=. 2024 , eprint=

    2024

  61. [61]

    2018 , eprint=

    Shampoo: Preconditioned Stochastic Tensor Optimization , author=. 2018 , eprint=

    2018

  62. [62]

    2024 , eprint=

    A New Perspective on Shampoo's Preconditioner , author=. 2024 , eprint=

    2024