BERT_VDA

This repository contains a BERT encoder saved after eye-tracking fine-tuning in the VDA_ET workflow.

The temporary token-level regression head used during training is not included. Load the checkpoint with AutoModel.from_pretrained for downstream encoder analysis or continued fine-tuning.

from transformers import AutoModel, AutoTokenizer

model_id = "calogero-jerik-scozzaro/BERT_VDA"
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True)
model = AutoModel.from_pretrained(model_id)

Training metadata

Field	Value
all_train_texts	lupo, lupoe, ciclisti, ciclistis
batch_size	8
epochs	100
learning_rate	2e-05
max_length	256
measures	FFD, FPRT, TFT, RRT, skipped, FPF, RR
num_train_sentences	36
source_model	dbmdz/bert-base-italian-uncased
stage	1
stage_train_texts	lupo, lupoe, ciclisti, ciclistis
test_texts	HumanRights, LearningMobility
variant	BERT_VDA

The uploaded files include et_label_scaler.json, which records the min-max scaling statistics used for the eye-tracking labels.