Samir Gadre

Samir Gadre, who publishes as Samir Yitzhak Gadre, is an AI researcher and an equal contributor to DataComp alongside Gabriel Ilharco and Alex Fang. His subsequent work includes language-model scaling studies and DataComp-LM, a testbed for training-data experiments. As of September 2026, his personal website describes him as a member of Anthropic’s technical staff on the pre-training team.[2][4][5][6]

Gadre, Ilharco, Fang and their collaborators introduced DataComp in 2023 to give multimodal dataset design the experimental attention usually devoted to architectures and training algorithms. Participants filter a candidate pool of 12.8 billion image-text pairs from Common Crawl or curate new data sources, then run standardized CLIP training and evaluate the resulting models on 38 downstream test sets. The shared training and evaluation procedure lets researchers compare dataset choices across compute scales spanning four orders of magnitude.[3][4]

The collaboration reported that its selected DataComp-1B dataset enabled a CLIP ViT-L/14 trained from scratch to reach 79.2% zero-shot ImageNet accuracy—3.7 percentage points above OpenAI’s CLIP ViT-L/14 using the same training procedure and compute. This result concerned the selected training set, distinct from the much larger candidate pool.[3]

Gadre appeared in the DataComp segment of Latent Space’s December 2023 NeurIPS recap, which also included a DataComp poster-session segment with him and Alex Dimakis.[1]

Gadre was the first-listed author of Language models scale reliably with over-training and on downstream tasks, initially submitted in March 2024 and listed at ICLR 2025. The study addressed two gaps between conventional scaling studies and practical model development: models are often trained beyond compute optimality to reduce inference costs, and their usefulness is judged on downstream tasks rather than next-token loss alone. Gadre and his collaborators built a testbed of 104 models, ranging from 0.011 billion to 6.9 billion parameters, trained on three data distributions.[5][2]

The team fitted scaling laws that extrapolated across model size and the amount of over-training, and proposed a power law relating perplexity to downstream performance. They reported predicting validation loss for a 1.4-billion-parameter model trained on 900 billion tokens and a 6.9-billion-parameter model trained on 138 billion tokens from experiments requiring 300 times less compute. Predictions of average downstream top-1 error used experiments requiring 20 times less compute. These ratios describe the predictive experiments, rather than reductions in the cost of training the final models.[5]

Gadre also coauthored DataComp-LM, initially submitted in June 2024 and listed at NeurIPS 2024. It extended controlled dataset experimentation to language models with a 240-trillion-token Common Crawl corpus, OpenLM-based pretraining recipes and 53 downstream evaluations. Participants can test deduplication, filtering and data mixing at model sizes from 412 million to 7 billion parameters. The collaboration found model-based filtering important for training-set quality; its revised abstract reported that DCLM-Baseline enabled a 7-billion-parameter model trained from scratch on 2.6 trillion tokens to reach 64% five-shot MMLU accuracy.[6][2]

Before the dataset and language-model studies, Gadre worked on embodied perception and interaction. He describes his earlier career as a Columbia PhD student advised by Shuran Song, also working with Ludwig Schmidt.[2][7]

In the 2021 paper Act the Part, Gadre, Kiana Ehsani and Song developed a method for discovering and segmenting articulated object parts through interaction. It couples action selection with motion segmentation, allowing an agent to expose object structure without semantic labels. The authors reported efficient part-discovery strategies, generalization to unseen categories and transfer from simulation to real-world data without fine-tuning.[7]