Datology

Datology, also branded DatologyAI, develops automated training-data curation: selecting, balancing, augmenting and ordering the data used to train AI models. Its platform manages the path from data in blob storage to training dataloaders, with deployment on customer infrastructure. The company’s work connects data research to model development, including pretraining datasets for Arcee AI and domain-specific mid-training data for Thomson Reuters.[3][8][12]

Its co-founders are Ari Morcos, CEO; Bogdan Gaza, CTO; and Matthew Leavitt. Morcos previously worked at Meta’s FAIR and DeepMind, Gaza brought infrastructure experience from Amazon and Twitter and was a Moonsense co-founder and CTO, and Leavitt previously headed data research at MosaicML and worked at FAIR. Datology’s financing and public product introduction preceded the research releases and customer collaborations that followed.[2][3][4]

Datology raised an $11.65 million seed round in September 2023 and publicly announced it with its February 21, 2024 introduction. Amplify Partners’ Sarah Catanzaro and Mike Dauber led the round, with participation from Radical Ventures, Conviction Capital, Outset Capital and Quiet Capital. At the introduction, Datology was working with a select group of customers and planned a broader release later that year.[3][4]

The initial service went beyond removing bad examples. It selected informative data in the context of the whole dataset, adjusted redundancy according to the complexity of each concept, rebalanced long-tailed distributions, applied targeted augmentation and optimized ordering and batch construction. Datology described deployment on premises or in a customer VPC and claimed petabyte-scale support across text, images, video, audio, tabular, genomic and geospatial data.[3]

On May 7, 2024, Morcos announced a $46 million Series A led by Felicis Ventures’ Viv Faga and Astasia Myers. Participants included Radical Ventures, Amplify Partners, Elad Gil, M12 and the Amazon Alexa Fund. Datology reported more than $57.5 million in total capital raised at that point; Morcos said the funding would expand the team and research compute across modalities.[4]

Released as a research report on August 18, 2025, BeyondWeb is the synthetic-data component of Datology’s pretraining curation pipeline. It rephrases existing documents into varied formats and styles—for example, question–answer pairs or passages with a more pedagogical tone—and restructures content to improve information density. Pratyush Maini and Vineeth Dorna led the core research. The approach draws knowledge from source documents while using a model to transform how that knowledge is presented.[5]

Datology tested models from 1 billion to 8 billion parameters, with training budgets reaching 1 trillion tokens. In its reported 8B experiment, BeyondWeb reached RedPajama’s 180B-token baseline performance after 23.2B tokens and Nemotron-Synth’s after 66.2B tokens: 7.7× and 2.7× fewer training tokens, respectively. These comparisons averaged zero-shot and five-shot accuracy across fourteen tasks, using 60% RedPajama and 40% synthetic tokens for the synthetic-data comparisons. The ratios measure convergence to baseline accuracy and exclude synthetic-generation cost. At the same 180B-token budget, a 3B BeyondWeb model also outperformed an 8B Cosmopedia model.[5]

Synthetic generation also prompted an infrastructure change. Datology replaced Slurm on an AWS Hyperpod H100 cluster with Ray and vLLM on Kubernetes after finding the original setup rigid, slow to iterate on and difficult to track experimentally. The replacement integrated with its Ray and Spark curation pipeline and supported heterogeneous customer clusters and GPU types. Core infrastructure contributors were Parth Doshi, Aldo Carranza, Fan Pan, Jack Urbanek and Paul Burstein.[5]

Arcee partnered with Datology as it moved from fine-tuning existing models to pretraining its own foundation models. Datology supplied quality filtering, embedding-driven selection, distribution balancing, source mixing and synthetic augmentation. Arcee progressed from its dense AFM-4.5B in July 2025 to Trinity Nano and Mini in December 2025 and then Trinity-Large-Thinking. For the larger-model project, Datology reported over 8 trillion synthetic tokens paired with 12 trillion curated web tokens, with synthesis clusters peaking at 2,048 H100 GPUs. Arcee handled model development, training infrastructure and post-training.[8]

The ÜberWeb report, published February 18, 2026, examined curation for thirteen non-English languages. Datology adapted filtering, embedding selection, synthetic rephrasing and mixing to each language’s scripts, distributions and data availability. In controlled bilingual experiments, it reported that curating English improved non-English performance in twelve of thirteen languages, averaging a 3.91% relative gain; tailored curation of both languages produced substantially larger gains. Here, “uncurated” baselines were random samples from DCLM and FineWeb2 without additional Datology curation.[7]

ÜberWeb described a broader 20-trillion-token corpus and experiments training 3B and 8B models on a random 1T-token subset. The curriculum increased multilingual content from 5% to 10% to 20% across three phases, yielding a 7.75% multilingual allocation overall. Trinity Large Base was pretrained on 17T tokens drawn from the broader corpus. Core contributors were Aldo Gael Carranza, Kaleigh Mentzer and Ricardo Pio Monti.[7]

Datology released the Luxical library and Luxical-One model under Apache 2.0 on December 12, 2025. Luxical prioritizes throughput, long documents and coarse document-to-document relationships for clustering, classification and semantic deduplication. It combines bag-of-ngrams features and TF-IDF scaling with an efficient sparse projection and nonlinear dense layers. Training distills a teacher model’s document-similarity distributions, allowing compact embeddings without requiring labeled examples from the target domain.[6]

Luxical-One produces 192-dimensional English-language embeddings. It uses the BERT uncased tokenizer and two million common FineWeb 5-grams, and was trained for three epochs on fifty million FineWeb documents with snowflake-arctic-embed-m-v2.0 as its teacher. Datology reported that it had already moved multiple curation workflows to Luxical. Its design serves large-scale document organization rather than optimizing every embedding task or fine-grained retrieval ranking.[6]

In 20/20 Vision Language Models, submitted May 12, 2026, Datology studied curation of the MAmmoTH-VL single-image subset while holding architecture, training recipe and compute fixed. The authors reported an average improvement of 11.7 percentage points across twenty public VLM benchmarks and 11.3 points across nine capability axes of DatBench, Datology’s evaluation suite. The work credits Siddharth Joshi, Haoli Yin, Rishabh Adiga and thirty other authors.[9]

The June 25 follow-up, Brevity is the Soul of Inference Efficiency, examined how curated training data changed response length and compute per correct answer. Using matched curated and baseline models at 1B, 2B and 4B, Datology reported 1.6×–2.3× lower estimated FLOPs per correct answer alongside 11.9–14.2 percentage-point accuracy gains. Its decode-cost proxy multiplies active parameters by output length and omits attention costs. On the captioning subset, fewer total hallucinated mentions were explained by shorter captions, rather than improved grounding per character. Matthew Leavitt and Sid Joshi were the core contributors.[10]

In August 2026, Datology introduced DataSmith, an autonomous research harness built on Claude Code and its curation environment. A persistent loop agent coordinates strategist, scout, diagnosis and synthesis agents, while an experiment tree records data recipes, results and alternative research directions. Data preparation and training can run asynchronously, with up to three concurrent training jobs. The system proposes interventions, prepares and decontaminates data, trains models, diagnoses failures and chooses subsequent experiments.[11]

Datology evaluated DataSmith using CurateBench, which fixes the base model, optimizer, SFT recipe and evaluations while allowing the agent to change training data. Each submission trains for 367 million tokens on eight H100s, with up to twenty submissions per session. Across three base models and four harness LLMs, Datology reported that DataSmith beat Claude Code in all twelve matched comparisons by an average 5.1 percentage points on AIME2025, GSM8K, HumanEval and BFCL. Aochong Oliver Li and Haoli Yin received equal-contribution credit.[11]

Thomson Reuters and Datology ran an early-2026 proof of concept applying mid-training data curation to proprietary legal data. Thomson Reuters subsequently launched Thomson on August 24, 2026, developing the model in-house with Imperial College London, DatologyAI and Lambda as technical partners. It continued training open-weight foundations through domain-specific mid-training, preference optimization and reinforcement learning. Datology supported mid-training data curation; the models and overall results belonged to Thomson Reuters’ full development effort.[12]

The collaboration curated a 200B-token mid-training dataset from a combined public and proprietary source pool exceeding 19T tokens, including more than 2T tokens of Thomson Reuters content. The mixture was roughly thirds curated proprietary documents, synthetic rephrasings using an adaptation of BeyondWeb, and curated general-capability replay data to preserve and strengthen existing skills. The pipeline ran entirely inside Thomson Reuters’ cloud account. This extended Datology’s document-rephrasing and curation methods from general pretraining into specialized model development on private data.[12]