Hypernova-60B
A compressed language model for agentic workflows and tool use.
My contribution Contributed to model development, optimized inference and distillation pipelines, and built synthetic training-data pipelines.
Models & code
Open models, research software, and personal projects, with my contribution to each model.
FLUX.1 Krea
An open image-generation model with a focus on aesthetics and photorealism.
My contribution Contributed to model development at Krea and optimized image and video fine-tuning pipelines with distributed training, including FSDP2.
GoLLIE
A language model that follows annotation guidelines to extract information from text.
My contribution Co-developed GoLLIE as part of my research on language models and information extraction at HiTZ.
Latxa
Instruction-tuned language models for Basque, evaluated on language understanding and generation.
My contribution Trained Latxa on the Leonardo supercomputer, using up to 512 GPUs across 128 nodes.
Medical-mT5
An open multilingual text-to-text language model for the medical domain.
My contribution Contributed to model development and co-authored the MedicalMT5 paper as first author.
Veridika.ai
An AI agent framework for real-time fact-checking.
My contribution Building the agent framework and its online demo.
Projects & Tools
AI-Generated GTAV
A deep learning project that uses Diffusion Transformers (DiT) to generate Grand Theft Auto V driving footage.
NoticIA
LLM finetuning and evaluation library for the NoticIA dataset of 850 Spanish news with human-written summaries.
Clickbait Fighter
AI that generates one-sentence summaries of clickbait news articles. Trained on 8×A100; deployed with vLLM and Ray.
Sequence Labeling with LLMs
Sequence labelling with LLMs via Text2Text constrained generation, built on Transformers + Accelerate.
T-Projection
High-quality annotation projection for sequence labeling datasets, built on Transformers + Accelerate.
LM Contamination Index
Manually curated database of contamination evidence for language models.
Context-enriched multilingual NER
Candidate generation + knowledge-base linking + fine-grained classification using retrieved knowledge.
Easy-Translate
Translate large text files with a single command. Easy for beginners, customizable for power users.
Easy Label Projection
Project labels across datasets using mGiza, FastAlign, SimAlign or AWESOME to generate resources for low-resource languages.
MetaVec
Monolingual and cross-lingual meta-embedding generation and evaluation framework.
Self Driving Car in Video Games
Supervised deep network that learns to drive in GTA V from human-labelled data.