Robin Armingaud

Robin Armingaud

Research engineer at CEA List

Robin Armingaud

I am a research engineer at CEA List, working on large language models (LLMs) for manufacturing applications. My research focuses on adapting LLMs for domain-specific tasks with an emphasis on efficiency, as well as exploring few-shot and zero-shot learning for information extraction. I enjoy tackling real-world problems with LLMs, and I regularly participate in Kaggle competitions and conference workshop.

News

  • [2025] GLiDRE, our new document-level relation extraction model for zero-shot relation extraction, will be released soon! Inspired by the excellent GLiNER, it’s designed for low-data scenarios.

  • [2025] We participated again in the EvaLLM 2025 workshop. Our participation paper will be available shortly!

  • [2025] My article on the TextMine 2025 competition is now online! Read it here: GLiDRE

  • [2025] The first (buggy!) version of GLiDRE ranked 10th in the TextMine 2025 Kaggle competition. Promising results.

  • [2025] Participation in the Drawing with LLMs Kaggle competition. My solution involved fine-tuning a Mistral-7B model using synthetic data generated with Gemini.

  • [2025] Participation in the AI Mathematical Olympiad – Progress Prize 2 Kaggle competition. We explored different prompt strategies using DeepSeek R1.

  • [2024] We successfully adapted a RoBERTa model for the manufacturing domain using curated internet data and advanced data selection strategies, achieving state-of-the-art results on internal benchmarks.

  • [2024] We won 1st place at the EvaLLM 2024 challenge, which required extracting entities from just four training examples. Our winning approach combined GLiNER, synthetic data generation with a Mistral model, and majority voting for robust performance. Read it here: CEA-List@EvalLLM2024

  • [2024] Participation in the LLM Prompt Recovery Kaggle competition. Our solution combined DeBERTa classifiers for template filling, a fine-tuned T5 model trained on synthetic data, and a prompted Mistral-7B model to generate candidate prompts. A final judge model selected the best candidate.