A risk classifier for minors that fits in a phone
Detecting grooming, harassment or threats in a minor's conversations without the conversation leaving the phone. The line generates half a million synthetic examples, fine-tunes a …
Measuring quantized LLMs: what is lost when an open model is compressed
Almost 300 GGUF files from dozens of families, served with llama.cpp and scored on function calling, code and long context, with speed and energy per card. The line that reserves the most …
A council of small models against a 120B one: automated jailbreaking and who judges it
Several small open models deliberate to build prompts that make gpt-oss-120b give in. The line measures how much deliberating adds over attacking alone, and devotes as much effort to the …
Self-alignment: how much safety to teach a language model without making it useless
A 3-billion-parameter Llama 3.2, LoRA fine-tuning and one question: what dose of safety data is needed for it to refuse what is harmful without refusing everything. The line tries fixed …
What really runs on our cluster, read from the scripts themselves
Four months of the compute cluster of the ANTS group at the University of Murcia, reconstructed from the Slurm jobs themselves and anonymised: what is really computed and what engineering is …