Loubna Ben Allal

Research

Loubna Ben Allal began at Hugging Face working on code generation as part of the core BigCode team behind The Stack, StarCoder and StarCoder2. The SantaCoder paper, which Ben Allal first-authored in 2023, received the best paper award at the ICLR 2023 Deep Learning for Code workshop.¹⁶

Ben Allal then turned to small language models, leading the SmolLM, SmolLM2 and SmolLM3 releases and building pretraining datasets such as FineWeb-Edu. The SmolLM2 paper, a spotlight at COLM 2025, describes overtraining a 1.7-billion-parameter model on a multi-stage mix of web, code, math and instruction data, and Ben Allal also wrote The Smol Training Playbook, a guide to building small language models.¹²³

More recently Ben Allal has worked on AI for biology. Carbon, an open generative DNA foundation model, was described in a May 2026 bioRxiv preprint with Ben Allal as first author.¹⁷

Talks

Ben Allal has spoken about model training at venues including Stanford's CS25 seminar in May 2024, the LLM4Code workshop at ICSE 2024, the NLP Summit 2024, and GPU MODE in January 2026, where the talk presented The Smol Training Playbook.¹

Work and contributions

• SmolLM

  Leads the SmolLM, SmolLM2, and SmolLM3 releases, according to Ben Allal's site, and is first author of the SmolLM2 paper on data-centric training of a small language model. The project repository describes compact open models meant to run on-device and includes local inference guides.¹²³

• SmolLM3-3B

  Part of the SmolLM3 work Ben Allal leads; the Apache-2.0 model card describes a 3B-parameter open model with quantized checkpoints for local runtimes.¹⁴

• FineWeb Open data

  Co-author of the paper introducing the FineWeb datasets, listed third.⁸

• StarCoder2-15B Open weights

  Third-listed author of the paper introducing StarCoder2 and The Stack v2, which the StarCoder2-15B model card cites.⁹

• StarCoder2-3B Open weights

  Third-listed author of the paper introducing StarCoder2 and The Stack v2, which the StarCoder2-3B model card cites.⁹

Timeline

1. 2022Co-authored The Stack, a dataset of permissively licensed source code for BigCode.¹
2. Jan 2023Posted SantaCoder as first author.⁶
3. 2024Co-authored the StarCoder2 and The Stack v2 paper.¹
4. 2024Co-authored the FineWeb datasets paper, a spotlight at NeurIPS 2024.¹
5. Feb 2025Posted the SmolLM2 paper as first author.²
6. Jan 2026Presented The Smol Training Playbook at GPU MODE.¹
7. May 25, 2026Posted the Carbon generative DNA foundation model preprint on bioRxiv as first author.⁷

Honors and recognition

• Best Paper, ICLR 2023 Deep Learning for Code (DL4C) Workshop, for SantaCoder · ICLR 2023 DL4C Workshop, 2023¹

Where they work

• Hugging Face 34 people in the wiki

Related people

• Guilherme PenedoAlso at Hugging Face
• Hynek KydlíčekAlso at Hugging Face
• Anton LozhkovAlso at Hugging Face
• Yacine JerniteAlso at Hugging Face
• Edward BeechingAlso at Hugging Face
• Vaishaal ShankarDatasets and data · Language models
• Sheshansh AgrawalLanguage models · Datasets and data
• Aran KomatsuzakiLanguage models · Datasets and data
• Avi SilLanguage models · Datasets and data
• Adina WilliamsLanguage models · Datasets and data