Dolma vs The Stack v2

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.

Aggregates 67 TB of source code from the Software Heritage archive, filtering for permissively licensed repositories (MIT, Apache 2.0, BSD, etc.) across 600+ programming languages. Uses automated license detection and validation to ensure legal compliance for model training. Implements a rigorous deduplication pipeline at file and repository levels to eliminate redundant training data and reduce dataset bloat.

Unique: Largest open-source code dataset at 67 TB with automated opt-out governance allowing repository owners to request removal, combined with rigorous deduplication and PII removal pipeline — no other public dataset offers this scale with legal compliance and community control mechanisms

vs alternatives: Larger and more legally compliant than GitHub's CodeSearchNet (14M files) or Google's BigQuery public datasets, with explicit opt-out governance vs. implicit inclusion, and covers 600+ languages vs. Codex training data's undisclosed language distribution

Implements a community-driven opt-out system where repository owners can request removal of their code from the dataset without legal takedown notices. Maintains a registry of excluded repositories and re-applies exclusions during dataset updates. Provides transparent governance documentation and a clear submission process for removal requests, balancing open access with creator rights.

Unique: First large-scale code dataset to implement opt-out governance at dataset level rather than relying solely on license compliance, with transparent registry and community submission process — shifts power from dataset creators to code contributors

vs alternatives: More respectful of creator autonomy than GitHub Copilot's training approach (no opt-out) or academic datasets (one-time snapshot), and more scalable than individual DMCA takedowns

Automated pipeline that scans source code for personally identifiable information (email addresses, API keys, SSH keys, credit card patterns, phone numbers) and removes or redacts them before dataset release. Uses regex patterns, entropy-based detection for secrets, and heuristic rules to identify sensitive data. Operates at file level with configurable sensitivity thresholds to balance data utility against privacy risk.

Unique: Combines regex pattern matching, entropy-based secret detection, and heuristic rules in a unified pipeline with configurable sensitivity — more comprehensive than simple regex-only approaches, but trades off false positive rate against security coverage

vs alternatives: More thorough than GitHub's secret scanning (which only flags known patterns) because it includes entropy-based detection for unknown secret formats, but less accurate than specialized tools like TruffleHog due to language-agnostic approach

Indexes 67 TB of source code across 600+ programming languages with language-aware metadata (syntax, file extension, language family). Enables retrieval by language, license, repository, or code patterns. Uses Software Heritage's existing indexing infrastructure as foundation, augmented with language detection and classification. Supports both bulk download and filtered queries for specific language subsets.

Unique: Leverages Software Heritage's existing language detection and indexing infrastructure, then augments with BigCode-specific language classification and filtering — avoids reinventing language detection while providing dataset-specific query capabilities

vs alternatives: More comprehensive language coverage (600+ languages) than GitHub's Linguist (500+ languages) and more accessible than Software Heritage's raw API because it's pre-filtered for permissive licenses and deduplicated

Removes duplicate code files and repositories using content hashing (SHA-256 or similar) and fuzzy matching for near-duplicates. Operates in two stages: exact deduplication via hash matching, then fuzzy matching (e.g., Jaccard similarity or MinHash) to catch semantically identical code with minor formatting differences. Preserves one canonical copy of each unique code pattern while removing redundant training examples.

Unique: Two-stage deduplication combining exact hash matching with fuzzy similarity matching (likely MinHash or Jaccard) to catch both identical and near-identical code — more thorough than single-stage approaches but computationally expensive

vs alternatives: More aggressive deduplication than CodeSearchNet (which uses simple hash matching) because it catches near-duplicates, but less semantic than clone detection tools (which understand code structure) because it's content-based

Integrates with Software Heritage's comprehensive archive of 200+ million repositories and their full version control history. Extracts source code snapshots from Software Heritage's Git/Mercurial/SVN repositories, preserving repository metadata (commit history, author info, timestamps). Provides access to code at specific points in time, enabling historical analysis or training on code evolution patterns.

Unique: Leverages Software Heritage's universal code archive (200M+ repositories) as data source, providing access to code that would be impossible to collect via GitHub API alone — enables training on archived/deleted repositories and non-GitHub platforms (GitLab, Gitea, etc.)

vs alternatives: More comprehensive than GitHub-only datasets because it includes code from GitLab, Gitea, SourceForge, and other platforms archived by Software Heritage; more legally defensible than web scraping because it uses an established, community-maintained archive

Tracks and validates SPDX license identifiers for each repository, ensuring only permissively licensed code (MIT, Apache 2.0, BSD, etc.) is included. Maintains license metadata alongside code files, enabling downstream users to verify legal compliance. Implements license hierarchy and compatibility checking to handle dual-licensed or complex licensing scenarios.

Unique: Combines automated SPDX detection with manual review and maintains license metadata alongside code, enabling downstream users to verify compliance — more transparent than datasets that simply claim 'permissive licenses' without proof

vs alternatives: More legally rigorous than GitHub's CodeSearchNet (which doesn't validate licenses) and more transparent than Codex training data (which doesn't disclose license filtering at all)

Maintains versioned snapshots of the dataset (e.g., v2.0, v2.1) with documented changes between versions (new repositories added, deduplication improvements, PII removal updates). Provides checksums and manifests for reproducibility, enabling researchers to cite specific dataset versions and reproduce results. Tracks dataset lineage and transformation history.

Unique: Maintains semantic versioning and detailed changelogs for dataset releases, enabling researchers to cite specific versions and understand dataset evolution — more rigorous than one-off dataset releases without versioning

vs alternatives: More reproducible than academic datasets that are released once without versioning, and more transparent than commercial datasets (Codex) that don't disclose version history or changes