Dolma vs The Pile

Dolma aggregates 3 trillion tokens from 7 heterogeneous sources (Common Crawl, The Stack, peS2o, Project Gutenberg, Wikipedia, Wikibooks, C4) with fully documented filtering criteria, deduplication methods, and mixing ratios. The composition system enables researchers to understand exactly which data proportions and quality thresholds were applied, making training runs reproducible across different teams and hardware configurations. Data is segmented into pretraining, mid-training, and post-training pools to support staged model development.

Unique: Dolma's distinguishing feature is comprehensive documentation of data curation decisions (exact filtering rules, deduplication methods via Duplodocus, mixing ratios) released alongside trained models (OLMo 7B, 32B), enabling full reproducibility. Most pretraining datasets (C4, The Pile, ROOTS) document composition at a high level but not the specific algorithmic rules applied. Dolma's integration with OlmoTrace enables tracing model outputs back to source training documents, providing data provenance that most datasets lack.

vs alternatives: Dolma provides greater transparency and reproducibility than C4 or The Pile through documented filtering rules and deduplication specifications, while offering more diverse source coverage (code + academic + literary) than web-only datasets like C4, though it is smaller than ROOTS (1.6T vs 3T tokens) and less frequently updated than continuously-refreshed web crawl datasets.

Dolma implements source-specific filtering pipelines using documented rules applied through tools like Datamap-rs (large-scale data cleaning) and Duplodocus (fuzzy deduplication). Each of the 7 sources undergoes tailored quality filtering appropriate to its characteristics: web crawl data is filtered for language and content quality, code is filtered for license and syntax validity, academic papers are filtered by venue quality, and literary text is filtered for encoding and completeness. Filtering rules are explicitly documented to enable researchers to understand and potentially modify quality thresholds.

Unique: Dolma's filtering approach is distinguished by source-specific quality criteria (e.g., academic papers filtered by venue quality, code filtered by license validity) rather than uniform filtering across all data. The integration of Duplodocus for fuzzy deduplication (vs. exact-match deduplication) is more sophisticated than simple hash-based approaches, enabling detection of near-duplicate content across sources. Documentation of exact filtering rules is rare in published datasets.

vs alternatives: Dolma's documented, source-specific filtering is more transparent than C4's undisclosed filtering rules, and more sophisticated than The Pile's simple language detection, though it requires external tools (Datamap-rs, Duplodocus) rather than providing integrated filtering infrastructure like some commercial training platforms.

Dolma's post-training data pool is designed for use with Open Instruct, Allen AI's instruction tuning framework, enabling seamless transition from pretraining to instruction tuning. The post-training pool contains instruction-formatted data (format unspecified) optimized for alignment and capability refinement. Integration with Open Instruct provides data loading, instruction formatting, and training orchestration for the post-training phase. This integration enables researchers to implement the full training pipeline (pretraining → continued pretraining → instruction tuning) using coordinated Dolma and Open Instruct components.

Unique: Dolma's post-training data pool with Open Instruct integration provides a coordinated instruction tuning solution that is rare in open-source ecosystems. Most datasets provide pretraining data only; Dolma's inclusion of post-training data and integration with Open Instruct enables end-to-end training without external instruction data curation. The simultaneous release of Dolma, OlmoCore, and Open Instruct provides a complete, reproducible training pipeline.

vs alternatives: Dolma's integrated post-training pipeline is more complete than datasets providing pretraining data only, though it is less flexible than using generic instruction datasets (e.g., Alpaca, ShareGPT) that support multiple training frameworks.

Dolma provides three distinct data pools optimized for different training stages: a pretraining pool for initial model training on diverse, general-purpose text; a mid-training pool for continued pretraining with potentially different source ratios or quality thresholds; and a post-training pool for instruction tuning and alignment. This segmentation enables researchers to apply different data compositions at different training phases without managing separate datasets, and allows for staged training strategies where model behavior is refined through targeted data exposure.

Unique: Dolma's segmentation into three explicit training phases (pretraining, mid-training, post-training) with separate downloadable pools is uncommon in published datasets. Most datasets provide a single corpus; Dolma's phase-specific segmentation enables researchers to implement sophisticated multi-stage training strategies without custom data partitioning. The integration with Open Instruct for post-training suggests end-to-end training pipeline support.

vs alternatives: Dolma's staged data segmentation is more structured than generic datasets like C4 or The Pile, which provide single corpora; it is comparable to commercial training platforms that offer phase-specific data curation, but with full transparency and reproducibility.

Dolma integrates with the OlmoTrace tool, which enables researchers to trace model outputs and behaviors back to the specific source documents in the training dataset that contributed to those outputs. This capability works by maintaining mappings between training data and model internals, allowing queries like 'which documents influenced this model's response?' or 'what is the source distribution of training data for this capability?'. Traceability is implemented through document-level tracking during preprocessing and training, enabling post-hoc analysis of model behavior in terms of training data composition.

Unique: OlmoTrace's document-level provenance tracing from model outputs back to training data is a rare capability in open-source LLM ecosystems. Most models provide no tracing mechanism; some provide source-level statistics but not output-specific tracing. Dolma's integration of traceability at the dataset level (maintaining document identifiers through preprocessing) enables this capability without post-hoc model modification.

vs alternatives: Dolma's provenance tracing via OlmoTrace provides transparency unavailable in most open models (which provide no tracing) and exceeds the source-level statistics provided by some datasets like C4, though it is less detailed than commercial model cards that sometimes include data attribution.

Dolma incorporates The Stack, a large-scale source code dataset, with code-specific filtering and quality control. Code data is filtered for license compliance (removing GPL and other restrictive licenses), syntax validity, and repository quality. The Stack integration provides access to diverse programming languages and coding patterns without requiring separate code dataset curation. Code is deduplicated using the same Duplodocus fuzzy deduplication as other sources, enabling detection of near-duplicate code across repositories.

Unique: Dolma's integration of The Stack with explicit license filtering (removing GPL) is distinctive because it enables commercial use of code-trained models while maintaining open-source compliance. Most code datasets (e.g., CodeParrot, GitHub Copilot training data) do not document license filtering or provide GPL-free variants. The combination of license filtering with fuzzy deduplication across code repositories is more sophisticated than simple exact-match deduplication.

vs alternatives: Dolma's code data provides license-compliant code training without GPL restrictions, making it suitable for commercial models, whereas The Pile and other generic datasets either include GPL code or lack code data entirely. However, it is smaller and less frequently updated than GitHub's full code index.

Dolma incorporates peS2o, a large-scale academic paper dataset, with venue-based quality filtering that prioritizes papers from high-impact conferences and journals. Academic papers are filtered by publication venue quality (e.g., top-tier conferences, high-impact journals) rather than citation count or other metrics, ensuring training data includes rigorous, peer-reviewed research. Paper text is extracted from PDFs and structured metadata, enabling models to learn from scientific writing and domain-specific knowledge. Academic data is deduplicated using the same fuzzy deduplication as other sources.

Unique: Dolma's use of venue-based quality filtering for academic papers (rather than citation count or other metrics) is distinctive because it prioritizes peer-review rigor over popularity, potentially reducing bias toward highly-cited but potentially flawed work. Integration of peS2o with explicit venue quality criteria is rare in published datasets; most datasets either exclude academic content or include it without quality filtering.

vs alternatives: Dolma's academic data provides peer-reviewed, venue-filtered content that exceeds generic datasets like C4 or The Pile in academic quality, though it is smaller and less frequently updated than full academic paper indices like arXiv or PubMed.

Dolma integrates web text from both Common Crawl (raw web crawl) and C4 (pre-filtered web text), with documented filtering rules for language detection, content quality, and toxicity. Web data undergoes source-specific filtering appropriate to its characteristics: Common Crawl data is filtered more aggressively due to lower baseline quality, while C4 data benefits from existing filtering. All web data is deduplicated using Duplodocus fuzzy deduplication to remove near-duplicate content across domains. The combination of two web sources with different filtering approaches provides diversity while maintaining quality standards.

Unique: Dolma's use of two complementary web sources (Common Crawl and C4) with source-specific filtering is distinctive because it balances raw coverage (Common Crawl) with pre-filtered quality (C4), providing diversity while maintaining standards. Most datasets use either raw crawls or pre-filtered sources, but not both. The documented filtering rules (though not detailed in available materials) enable reproducibility that most web datasets lack.

vs alternatives: Dolma's dual-source web data provides greater transparency and reproducibility than C4 alone, while offering broader coverage than C4-only datasets, though it is smaller and less frequently updated than continuously-refreshed web crawl datasets.

Combines 22 discrete, curated text datasets (academic papers, books, code, web text, specialized sources) into a single 825 GiB jsonlines corpus compressed with zstandard. The assembly approach prioritizes diversity across domains rather than size maximization, enabling language models trained on this corpus to develop broad cross-domain knowledge and generalization capabilities. Data is provided as-is without documented preprocessing, deduplication, or filtering pipelines, placing responsibility for data cleaning on downstream users.

Unique: Pioneered the multi-domain curation approach by intentionally combining 22 diverse, high-quality subsets (academic papers, books, code, web, specialized sources) rather than scraping a single massive web corpus. This architectural choice prioritizes knowledge breadth and domain coverage over raw scale, influencing the design of subsequent open datasets like LAION, RedPajama, and Falcon-Refinedweb.

vs alternatives: Broader domain coverage than Common Crawl-only datasets (e.g., C4) and higher quality than raw web scrapes due to curation of academic, code, and book sources; smaller than Falcon-Refinedweb (1.5T tokens) but more carefully curated and widely adopted as a benchmark for model evaluation

Provides a standardized evaluation metric (Pile Bits Per Byte, or BPB) that measures language model perplexity across the full 22-subset corpus, enabling comparison of model generalization across diverse text domains. The metric is computed by evaluating a trained model on held-out portions of each subset and aggregating results, producing a single scalar score where lower values indicate better cross-domain performance. This approach surfaces domain-specific weaknesses that single-domain metrics would miss.

Unique: Introduced BPB (Bits Per Byte) as a standardized metric for evaluating language model performance across a curated multi-domain corpus rather than a single domain or random web text. This approach surfaces generalization gaps that domain-specific metrics (e.g., code completion accuracy, translation BLEU) would miss, establishing a precedent for multi-domain evaluation in subsequent benchmarks (MMLU, HELM).

vs alternatives: More comprehensive than single-domain metrics (e.g., GLUE for NLU, HumanEval for code) because it evaluates across 22 domains simultaneously; more reproducible than web-scale benchmarks (e.g., zero-shot on random web text) due to fixed, curated evaluation set, though leaderboard adoption remains limited due to sparse published results

Provides training data in a model-agnostic jsonlines format that integrates with standard ML frameworks (PyTorch, TensorFlow, Hugging Face) without requiring custom preprocessing or format conversion. The jsonlines + zstandard approach enables seamless integration with existing dataloaders, tokenizers, and training pipelines, reducing friction for researchers adopting the dataset. No custom APIs or proprietary tools are required — standard open-source libraries suffice.

Unique: Uses standard, framework-agnostic jsonlines + zstandard format that integrates directly with PyTorch, TensorFlow, and Hugging Face without custom preprocessing or proprietary tools. This contrasts with proprietary formats (HDF5, custom binary formats) that require custom loaders, or single-framework datasets that lock users into specific ML libraries.

vs alternatives: More portable than proprietary formats because it uses standard jsonlines; more efficient than uncompressed text because zstandard compression reduces storage by ~3-4x; simpler than database formats (SQLite, Parquet) because jsonlines requires no schema definition or query language.

Encodes the 825 GiB corpus as jsonlines (one JSON object per line, typically with a 'text' field containing raw text) and compresses with zstandard (zstd), a modern compression algorithm offering faster decompression and better compression ratios than gzip. This format choice enables streaming decompression and line-by-line parsing without loading the entire dataset into memory, critical for training pipelines on resource-constrained hardware. The jsonlines structure allows metadata (e.g., source subset, document ID) to be stored alongside text.

Unique: Chose zstandard compression over gzip or bzip2, offering ~20% better compression ratios and 5-10x faster decompression speeds, critical for large-scale training pipelines where I/O is a bottleneck. Paired with jsonlines format to enable streaming decompression and line-by-line parsing without materializing the full 825 GiB dataset in memory.

vs alternatives: Faster decompression than gzip-compressed datasets (e.g., C4) and more memory-efficient than uncompressed datasets; jsonlines format is more flexible than binary formats (e.g., HDF5, TFRecord) for preserving metadata and enabling ad-hoc analysis, though slightly slower to parse than optimized binary formats

Explicitly enumerates the 22 constituent subsets of the Pile (academic papers from PubMed and ArXiv, books from Books3 and Gutenberg, code from GitHub, web text from OpenWebText2 and Pile-CC, specialized sources like USPTO patents, Ubuntu IRC, and Stack Exchange) and provides source attribution for each document. This transparency enables users to understand the composition of their training data, audit for potential biases or contamination, and selectively exclude subsets if needed. However, exact composition percentages and subset enumeration are not fully documented.

Unique: Pioneered explicit, multi-source composition transparency in large pretraining datasets by publicly naming 22 constituent subsets and their sources, establishing a precedent for data provenance documentation in subsequent datasets (RedPajama, Falcon-Refinedweb). This approach enables auditing and selective subset exclusion, though exact composition percentages remain undocumented.

vs alternatives: More transparent than Common Crawl-only datasets (e.g., C4) which provide minimal source attribution; comparable to RedPajama in subset enumeration but less detailed in per-document source labels and composition percentages

Includes curated subsets of academic papers (PubMed, ArXiv), specialized technical sources (USPTO patents, Stack Exchange), and code repositories (GitHub), providing dense coverage of high-signal, domain-specific text that is underrepresented in web-only corpora. These subsets are integrated into the broader corpus at a fixed ratio, ensuring that models trained on the Pile develop specialized knowledge in these domains without requiring separate fine-tuning. The inclusion of academic papers and code is particularly valuable for training models intended for scientific or technical applications.

Unique: Intentionally curated academic papers (PubMed, ArXiv) and code (GitHub) as core subsets rather than treating them as incidental web scrape byproducts, establishing a precedent for domain-specific data curation in pretraining. This approach ensures models trained on the Pile develop strong performance on technical and scientific tasks without requiring separate fine-tuning or domain-specific pretraining.

vs alternatives: More comprehensive academic and code coverage than web-only datasets (e.g., C4, Common Crawl); comparable to domain-specific datasets (e.g., CodeSearchNet for code, S2ORC for academic papers) but integrated into a single multi-domain corpus for broader generalization

Incorporates two book-focused subsets (Books3 and Gutenberg) providing long-form, narrative text with complex linguistic structures, enabling models to develop strong performance on coherent, multi-paragraph generation and understanding of narrative arcs. Books represent a fundamentally different text distribution than web text (longer documents, more complex grammar, narrative structure) and are valuable for training models intended for creative writing, summarization, or long-context understanding. The inclusion of both contemporary books (Books3) and public-domain classics (Gutenberg) provides temporal and stylistic diversity.

Unique: Explicitly includes book-focused subsets (Books3, Gutenberg) as core components rather than incidental web scrape byproducts, recognizing that long-form narrative text develops different linguistic capabilities than short web snippets. This architectural choice influences model performance on coherence, narrative structure, and long-context understanding.

vs alternatives: More comprehensive book coverage than web-only datasets (e.g., C4); comparable to book-specific datasets (e.g., BookCorpus) but integrated into a multi-domain corpus for broader generalization rather than domain-specific pretraining

Combines two web-derived subsets (OpenWebText2 and Pile-CC) providing broad coverage of diverse web text while applying quality filtering and deduplication to reduce noise compared to raw Common Crawl. OpenWebText2 is derived from URLs shared on Reddit (a proxy for human-curated quality), while Pile-CC is a filtered subset of Common Crawl. Together, these subsets provide web-scale coverage without the extreme noise and duplication of raw web scrapes, balancing breadth with quality.

Unique: Combines Reddit-curated web text (OpenWebText2) with filtered Common Crawl (Pile-CC) rather than relying on raw Common Crawl alone, applying implicit quality filtering through Reddit curation and explicit deduplication/filtering on Pile-CC. This hybrid approach balances web-scale coverage with quality, addressing a key limitation of earlier web-only datasets.

vs alternatives: Higher quality than raw Common Crawl (e.g., C4) due to Reddit curation and filtering; broader coverage than Reddit-only datasets; comparable to Falcon-Refinedweb in approach but with less documented filtering methodology