README.md

---
license: apache-2.0
task_categories:
- other
pretty_name: Olmix Swarm Datasets
size_categories:
- n<1K
tags:
- data-mixing
- language-models
- pretraining
- mixture-optimization
---

# Olmix Swarm Datasets

This repository contains proxy run swarm datasets for data mixing using **[Olmix](https://github.com/allenai/olmix/)**. Each swarm consists of multiple proxy training runs with different domain mixture ratios and their corresponding evaluation metrics across various downstream tasks. The swarm datasets here are based on **30M parameter** proxy models trained on **3B tokens**.

For more details, see our paper: [*Olmix: A Framework for Data Mixing Throughout LM Development*](https://arxiv.org/abs/2602.12237)

## Swarms

The dataset contains **36 swarms**. Each swarm consists of `ratios.csv`, `metrics.csv`, and `meta.json`.

```
dclm_swarm/                          # Main DCLM swarm (1 swarm)
├── study/
│   ├── rq2/                         # Swarm size experiments (3 swarms)
│   │   ├── dclm_swarm_18_topics/
│   │   ├── dclm_swarm_12_topics/
│   │   └── dclm_swarm_6_topics/
│   ├── rq3/                         # Swarm distribution experiments (5 swarms)
│   │   ├── dclm_dense_swarm/
│   │   ├── source_level_sparse_swarm/
│   │   ├── source_level_dense_swarm/
│   │   ├── dclm_swarm_strong_prior/
│   │   └── dclm_swarm_weak_prior/
│   └── rq6/                         # Constrained swarm for enforcing repetition (1 swarm)
│       └── dclm_constrained_swarm/
└── mixture_reuse/
    ├── real_world/
    │   ├── full_reuse/              # Full mixture reuse (12 swarms)
    │   │   ├── update1_add_stack_edu_seed{0,1,2}/
    │   │   ├── update2_add_more_sources_seed{0,1,2}/
    │   │   ├── update3_revise_pdfs_seed{0,1,2}/
    │   │   └── update5_partition_pdfs_seed{0,1,2}/
    │   ├── partial_reuse/           # Partial mixture reuse (6 swarms)
    │   │   ├── update1_add_stack_edu_recompute_software_dev_seed{0,1,2}/
    │   │   └── update5_partition_pdfs_recompute_mix_seed{0,1,2}/
    │   ├── swarm_reuse/             # Swarm reuse baseline (1 swarm)
    │   └── full_recomputation/      # Full recomputation baseline (1 swarm)
    └── theory_validation/           # Theoretical validation (6 swarms)
        ├── dclm_stackedu/
        ├── dclm_pdfs/
        ├── weak_mix_reuse_gap_dclm_stackedu/
        ├── intermediate_mix_reuse_gap_dclm_stackedu/
        ├── weak_mix_reuse_gap_dclm_pdfs/
        └── intermediate_mix_reuse_gap_dclm_pdfs/
```

- **`dclm_swarm/`** (1 swarm): 128 proxy runs over 24 [DCLM](https://arxiv.org/abs/2406.11794) topics (sparse); the primary swarm used throughout the paper.

- **`study/`** (9 swarms):
  - *RQ2 — Swarm Size* (3 swarms, Section 3.2): How many proxy runs are needed for m domains? Swarms over subsets of DCLM topics: 18 topics (129 runs total), 12 topics (130 runs total), and 6 topics (125 runs total).
  - *RQ3 — Swarm Distribution* (5 swarms, Section 3.3): Comparing swarm configurations over 24 DCLM topics or 7 sources, varying density (sparse vs. dense) and Dirichlet prior strength (weak vs. strong).
  - *RQ6 — Constrained Optimization* (1 swarm, Section 3.6): 128 proxy runs satisfying repetition constraints for R=6T requested tokens and k=4 repetition factor.

- **`mixture_reuse/real_world/`** (20 swarms, Section 5.1, Figure 10):
  - *Full reuse* (12 swarms): The entire previous mixture is reused as a single virtual domain. 3 seeds each for 4 updates:
    - Update 1 (add Stack-Edu): swarms over DCLM (1 virtual domain) and [Stack-Edu](https://arxiv.org/abs/2502.02737)
    - Update 2 (add more sources): swarms over DCLM+Stack-Edu (1 virtual domain) and new sources
    - Update 3 (revise PDFs): swarms over revised [PDF](https://arxiv.org/abs/2512.13961) source and unchanged domains (1 virtual domain)
    - Update 5 (partition PDFs): swarms over partitioned PDF topics and unchanged domains (1 virtual domain)
  - *Partial reuse* (6 swarms): Only the unaffected portion of the mixture is reused; affected domains are recomputed. 3 seeds each for 2 updates:
    - Update 1 (add Stack-Edu, recompute software development): swarms over DCLM \ software development (1 virtual domain), software development, and Stack-Edu
    - Update 5 (partition PDFs, recompute mix): swarms over PDF topics and sources
  - *Swarm reuse* (1 swarm): Proxy runs accumulated from all swarms compatible with the final 64-domain set.
  - *Full recomputation* (1 swarm): 512 proxy runs over the final 64 domains after all 5 updates. Subsampled at various budgets to measure full recomputation's performance as a function of run count.

- **`mixture_reuse/theory_validation/`** (6 swarms, Section 5.2): Experiments validating mixture reuse theory.
  - *DCLM+StackEdu* and *DCLM+PDFs*: Swarms for the full recomputation baseline in two settings.
  - *Weak/Intermediate mix reuse gap (DCLM+StackEdu)*: Swarms over DCLM (reusing a weak or intermediate mix) and Stack-Edu.
  - *Weak/Intermediate mix reuse gap (DCLM+PDFs)*: Swarms over DCLM (reusing a weak or intermediate mix) and PDFs.

## Data Format

### ratios.csv

- `run`, `name`, `index`: run identifier, full run name, and sequential index
- One column per domain: mixture weight (proportion of training tokens) assigned to that domain. Domain names follow the format `source:topic` (e.g., `dclm:politics`, `s2pdf:science_tech`, `stack-edu:Python`), where the prefix is the data source and the suffix is the topic within that source. Single-part names (e.g., `wikipedia`, `arxiv`) are single-domain sources.

### metrics.csv

- `run`, `name`, `index`: same identifiers as in `ratios.csv`
- One column per evaluation task: the model's **bits-per-byte (BPB)** score on that task. Lower BPB is better. Tasks span a wide range of benchmarks including MMLU, ARC, coding (HumanEval, MBPP), math (Minerva), and reading comprehension.

### meta.json

Each `meta.json` includes `swarm_id`, `description`, `category`, `notes`, `files`, and two fields used when configuring a fit in the [Olmix](https://github.com/allenai/olmix) repository:

- **`relative_sizes`**: The natural distribution of the domains, used in the KL penalty.
- **`token_counts`**: The final token count for each domain when constructing the repetition constraint.

## Usage

The primary use of these datasets is as input to the [Olmix GitHub repository](https://github.com/allenai/olmix) for constructing a proposed mix. Download `ratios.csv`, `metrics.csv`, and `meta.json` from the desired swarm folder and point to them in your Olmix config `.yaml` file (see [example config](https://github.com/allenai/olmix/blob/main/configs/fits/dclm_baseline.yaml)).

The dataset is also accessible via the HuggingFace datasets library (`load_dataset("allenai/olmix")`) for exploration purposes.

## Resources

- **GitHub**: [https://github.com/allenai/olmix](https://github.com/allenai/olmix)
- **Paper**: [https://arxiv.org/abs/2602.12237](https://arxiv.org/abs/2602.12237)

## Citation

If you use these datasets in your research, please cite:

```bibtex
@article{chen2026olmix,
  title={Olmix: A Framework for Data Mixing Throughout LM Development},
  author={Chen, Mayee F and Murray, Tyler and Heineman, David and Jordan, Matt and Hajishirzi, Hannaneh and Re, Christopher and Soldaini, Luca and Lo, Kyle},
  journal={arXiv preprint arXiv:2602.12237},
  year={2026}
}
```

## License

This dataset is released under the Apache 2.0 License.