BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:Fine-Tuning Transformers for Medical Reasoning with LoRA and Huggi
 ng Face Trainer
DTSTART;VALUE=DATE-TIME:20260717T090000Z
DTEND;VALUE=DATE-TIME:20260717T110000Z
DTSTAMP;VALUE=DATE-TIME:20260702T234222Z
UID:indico-contribution-216-1162@events.grnet.gr
DESCRIPTION:Speakers: Roman  Dolgopolyi (GRNET)\nThis talk demonstrates a 
 practical\, end-to-end notebook for fine-tuning a reasoning-capable transf
 ormer model for medical question answering. Participants will learn how to
  use Hugging Face Datasets and Trainer to handle the training workflow\, f
 rom loading and cleaning data to tokenization\, checkpointing\, evaluation
 \, and inference. The session demonstrates parameter-efficient fine-tuning
  with LoRA\, showing how a 3B-class Mistral reasoning model can be adapted
  on a single 16 GB GPU by training only small adapter weights instead of t
 he full model. The notebook combines MedReason and medical-o1 reasoning da
 tasets into a unified question\, chain-of-thought\, and answer format\, th
 en trains and evaluates the model on a small demo subset. By the end\, att
 endees will understand the key engineering choices behind efficient LLM fi
 ne-tuning and see a side-by-side comparison of base and fine-tuned model b
 ehavior on medical reasoning tasks\, including practical notes on GPU setu
 p\, mixed precision\, and resource cleanup for reproducible classroom demo
 s.\n\nhttps://events.grnet.gr/event/216/contributions/1162/
LOCATION:
URL:https://events.grnet.gr/event/216/contributions/1162/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Efficient training\, fine-tuning and inference of large-scale ML m
 odels
DTSTART;VALUE=DATE-TIME:20260717T080000Z
DTEND;VALUE=DATE-TIME:20260717T090000Z
DTSTAMP;VALUE=DATE-TIME:20260702T234222Z
UID:indico-contribution-216-1161@events.grnet.gr
DESCRIPTION:Speakers: Constantine  Dovrolis (The Cyprus Institute)\nThis t
 alk presents model-centric methods for efficient generative AI. It explain
 s why training and inference of LLMs are computationally heavy\, then cove
 rs model compression methods such as quantization\, neural network pruning
 \, low-rank approximations\, and knowledge distillation. It also introduce
 s efficient pre-training with mixed-precision acceleration and PHEW\, para
 meter-efficient fine-tuning methods such as LLM-Adapters\, LLaMA-Adapter\,
  P-Tuning\, and LoraHub\, and efficient inference techniques including spe
 culative decoding and KV-cache optimization.\n\nhttps://events.grnet.gr/ev
 ent/216/contributions/1161/
LOCATION:
URL:https://events.grnet.gr/event/216/contributions/1161/
END:VEVENT
END:VCALENDAR