Olmix Swarm Datasets

This repository contains proxy run data for data mixing using Olmix. Each swarm consists of multiple proxy training runs with different domain mixture ratios and their corresponding evaluation metrics across various downstream tasks.

For more details, see our paper: Olmix: A Framework for Data Mixing Throughout LM Development

Dataset Overview

The dataset contains 32 swarms organized into four main categories:

• Main Swarm (dclm_swarm/): The primary DCLM swarm used throughout our paper
• Research Question Studies (study/): Swarms for RQ3 (dense vs sparse, different priors) and RQ6 (constrained optimization)
• Mixture Reuse (mixture_reuse/): Real-world update scenarios and theory validation experiments
• Each swarm includes ratios.csv, metrics.csv, and meta.json with complete metadata

Usage

Primary Use Case: Proposing a Mix using OlmixBase

The primary way to use these datasets is to download these CSVs (ratios.csv, metrics.csv together) directly and use them with the Olmix GitHub (example). This is the recommended workflow for data mixture optimization.

Alternative: HuggingFace Datasets Library

You can also load all swarms at once using the HuggingFace datasets library (useful for exploration):

from datasets import load_dataset

# Load the entire collection
dataset = load_dataset("allenai/olmix")

# Browse available swarms
for example in dataset["train"]:
    print(f"{example['description']} - {example['num_runs']} runs, {example['num_domains']} domains")

# Filter by category
mixture_reuse_swarms = [ex for ex in dataset["train"] if "mixture_reuse" in ex["category"]]
print(f"Found {len(mixture_reuse_swarms)} mixture reuse swarms")

Dataset Structure

dclm_swarm/                          # Main DCLM swarm (1 swarm)
│   ├── ratios.csv
│   ├── metrics.csv
│   └── meta.json
├── study/
│   ├── rq3/                         # Dense vs sparse comparison (5 swarms)
│   │   ├── dclm_dense_swarm/
│   │   ├── source_level_sparse_swarm/
│   │   ├── source_level_dense_swarm/
│   │   ├── dclm_swarm_strong_prior/
│   │   └── dclm_swarm_weak_prior/
│   └── rq6/                         # Constrained optimization (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 (1 swarm)
    │       └── update4_remove_algebraicstack/
    └── theory_validation/           # Theory 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/

Dataset Fields

Each example in the dataset contains:

Field	Type	Description
swarm_id	string	Unique swarm identifier
description	string	Human-readable swarm description
category	string	Category/folder path (e.g., "dclm_swarm", "mixture_reuse/real_world/full_reuse")
notes	string	Detailed description of the swarm configuration
dataset_path	string	Relative path to the dataset
ratios	string	JSON string of domain mixture ratios (pandas DataFrame)
metrics	string	JSON string of evaluation metrics (pandas DataFrame)
num_runs	int	Number of proxy training runs
num_domains	int	Number of domains in the mixture
num_metrics	int	Number of evaluation metrics

Swarm Categories

Main Swarm (1 swarm)

• DCLM main swarm: 128 proxy runs over 24 DCLM topics (sparse), used throughout our paper

RQ3: Dense vs Sparse Comparison (5 swarms)

• DCLM dense swarm: 128 proxy runs over 24 DCLM topics (dense)
• Source level sparse swarm: 64 proxy runs over 7 sources (sparse)
• Source level dense swarm: 64 proxy runs over 7 sources (dense)
• DCLM swarm strong prior: 128 proxy runs with strong Dirichlet prior
• DCLM swarm weak prior: 128 proxy runs with weak Dirichlet prior

RQ6: Constrained Optimization (1 swarm)

• DCLM constrained swarm: 128 proxy runs with R=6T requested tokens and k=4 repetition factor

Mixture Reuse: Real World - Full Reuse (12 swarms)

Experiments where the entire previous mixture is reused as a single virtual domain:

• Update 1 (add Stack-Edu): 3 seeds, swarms over DCLM (1 virtual domain) and Stack-Edu
• Update 2 (add more sources): 3 seeds, swarms over DCLM+Stack-Edu (1 virtual domain) and new sources
• Update 3 (revise PDFs): 3 seeds, swarms over new PDF source and unchanged domains (1 virtual domain)
• Update 5 (partition PDFs): 3 seeds, swarms over PDF topics and unchanged domains (1 virtual domain)

Mixture Reuse: Real World - Partial Reuse (6 swarms)

Experiments where only part of the mixture is reused, with affected domains recomputed:

• Update 1 (add Stack-Edu, recompute software development): 3 seeds, swarms over DCLM \ software development (1 virtual domain), software development, and Stack-Edu
• Update 5 (partition PDFs, recompute mix): 3 seeds, swarms over PDF topics and source-level

Mixture Reuse: Real World - Swarm Reuse (1 swarm)

• Update 4 (remove AlgebraicStack): Swarm after AlgebraicStack is removed is compatible with final domain set

Mixture Reuse: Theory Validation (6 swarms)

Experiments validating mixture reuse theory:

• DCLM+StackEdu: Swarm over DCLM+StackEdu (for full recomputation baseline)
• DCLM+PDFs: Swarm over DCLM+PDFs (for full recomputation baseline)
• Weak mix reuse gap DCLM+StackEdu: Swarm over DCLM (reuse weak mix) and Stack-Edu
• Intermediate mix reuse gap DCLM+StackEdu: Swarm over DCLM (reuse intermediate mix) and Stack-Edu
• Weak mix reuse gap DCLM+PDFs: Swarm over DCLM (reuse weak mix) and PDFs
• Intermediate mix reuse gap DCLM+PDFs: Swarm over DCLM (reuse intermediate mix) and PDFs

Example: Analyzing Domain Distributions

import pandas as pd
import matplotlib.pyplot as plt

# Load a swarm
ratios = pd.read_csv("dclm_swarm/ratios.csv")
metrics = pd.read_csv("dclm_swarm/metrics.csv")

# Get domain columns (excluding metadata columns)
domain_cols = [c for c in ratios.columns if c not in ['run', 'name', 'index']]
metric_cols = [c for c in metrics.columns if c not in ['run', 'name', 'index']]

print(f"Domains: {len(domain_cols)}")
print(f"Metrics: {len(metric_cols)}")

# Plot distribution of a domain
plt.figure(figsize=(10, 6))
plt.hist(ratios[domain_cols[0]], bins=30, alpha=0.7)
plt.xlabel(f"{domain_cols[0]} Ratio")
plt.ylabel("Frequency")
plt.title(f"Distribution of {domain_cols[0]} in DCLM Swarm")
plt.show()

# Analyze metric correlations
corr_matrix = metrics[metric_cols].corr()
print(f"Metric correlation matrix: {corr_matrix.shape}")

Resources

• GitHub: https://github.com/allenai/olmix
• Paper: https://arxiv.org/abs/2602.12237

Citation

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

@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.