PreDA-base (Prefix-Based Dream Reports Annotation)

This model is a fine-tuned version of google-t5/t5-base on the annotated Dreambank.net dataset.It achieves the following results on the evaluation set:

Intended uses & limitations

This model is designed for research purposes. See the disclaimer for more details.

Training procedure

The overall idea of our approach is to disentangle each dream report from its annotation as a whole and to create an augmented set of (dream report; single feature annotation). To make sure that, given the same report, the model would produce a specific HVDC feature, we simply append at the beginning of each report a string of the form ``HVDC-Feature:'', in a manner that closely mimics T5 task-specific prefix fine-tuning.

After this procedure to the original dataset (~1.8K) we obtain approximately 6.6K items. In the present study, we focused on a subset of six HVDC features: Characters, Activities, Emotion, Friendliness, Misfortune, and Good Fortune. This selection was made to exclude features that represented less than 10% of the total instances. Notably, Good Fortune would have been excluded under this criterion, but we intentionally retained this feature to control against potential memorisation effects and to provide a counterbalance to the Misfortune feature. After filtering out instances whose annotation feature is not one of the six selected features, we are left with ~5.3K dream reports. We then generate a random split of 80%-20% for the training (i.e., 4,311 reports) and testing (i.e. 1,078 reports) sets.

Training

Hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.001
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 10
• num_epochs: 20
• mixed_precision_training: Native AMP
• label_smoothing_factor: 0.1

Metrics

Training Loss	Epoch	Step	Validation Loss	Rouge1	Rouge2	Rougel	Rougelsum
1.9319	1.0	135	1.8550	0.5264	0.3732	0.5090	0.5092
1.8158	2.0	270	1.7623	0.6482	0.5020	0.6238	0.6242
1.7508	3.0	405	1.7072	0.7429	0.6136	0.7142	0.7141
1.6998	4.0	540	1.6532	0.7753	0.6726	0.7517	0.7518
2.6261	5.0	675	2.7877	0.6512	0.5178	0.6319	0.6327
2.6021	6.0	810	2.4693	0.7583	0.6397	0.7384	0.7383
2.4839	7.0	945	2.3952	0.7914	0.6872	0.7684	0.7682
2.4471	8.0	1080	2.3693	0.7851	0.6768	0.7622	0.7621
2.4298	9.0	1215	2.3654	0.7881	0.6811	0.7649	0.7647
2.3975	10.0	1350	2.4163	0.7884	0.6785	0.7662	0.7660
2.4715	11.0	1485	2.4415	0.7861	0.6745	0.7629	0.7628
2.4541	12.0	1620	2.4401	0.7851	0.6730	0.7621	0.7620
2.4057	13.0	1755	2.4395	0.7856	0.6735	0.7621	0.7619
2.4491	14.0	1890	2.4387	0.7863	0.6739	0.7627	0.7624
2.4713	15.0	2025	2.4379	0.7870	0.6755	0.7635	0.7634
2.4635	16.0	2160	2.4375	0.7865	0.6742	0.7630	0.7628
2.4693	17.0	2295	2.4364	0.7869	0.6749	0.7630	0.7627
2.2091	18.0	2430	2.4366	0.7859	0.6743	0.7629	0.7626
2.4518	19.0	2565	2.4362	0.7866	0.6753	0.7635	0.7635
2.4663	20.0	2700	2.4362	0.7866	0.6744	0.7633	0.7631

Evaluation

We selected the best model via validation loss. The table below reposts overall and feature-specific scores.

Feature	rouge1	rouge2	rougeL	rougeLsum
Activities	62.6	51	58.4	58.3
Characters	84.7	79.5	82.9	82.9
Emotion	79.9	72.1	78.7	78.9
Friendliness	70.3	57	66.3	66.3
Good Fortune	71.4	22.8	71.7	71.5
Misfortune	65.6	48.2	65.2	65.1
Overall	78.7	67.4	76.3	76.3

Disclaimer

Dream reports and their annotation have been used in clinical settings and applied for diagnostic purposes. This does not apply in any way to our experimental results and output. Our work aims to provide experimental evidence of the feasibility of using PLMs to support humans in annotating dream reports for research purposes, as well as detailing their strengths and limitations when approaching such a task.

Framework versions

• Transformers 4.44.2
• Pytorch 2.1.0+cu118
• Datasets 3.0.1
• Tokenizers 0.19.1

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_id = "jrc-ai/PreDA-base"
device = "cpu"
encoder_max_length = 100
decoder_max_length = 50

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

dream = "I was talking with my brother about my birthday dinner. I was feeling sad."
prefixes = ["Emotion", "Activities", "Characters"]
text_inputs = ["{} : {}".format(p, dream) for p in prefixes]

inputs = tokenizer(
    text_inputs,
    max_length=encoder_max_length,
    truncation=True,
    padding=True,
    return_tensors="pt"
)

output = model.generate(
    **inputs.to(device),
    do_sample=False,
    max_length=decoder_max_length,
)

for decode_dream in output:
    print(tokenizer.decode(decode_dream, skip_special_tokens=True))

Dual-Use Implication

Upon evaluation we identified no dual-use implication for the present model. The model parameters, including the weights are available under CC0 1.0 Public Domain Dedication.

Cite

If you use our models in your research, please cite us as:

@InProceedings{10.1007/978-3-032-21477-5_13,
author="Bertolini, Lorenzo
and Comte, Valentin
and Ceresa, Mario
and Consoli, Sergio",
editor="Nicosia, Giuseppe
and Ojha, Varun
and Giesselbach, Sven
and Pardalos, M. Panos
and Umeton, Renato
and Emanuele, La Malfa
and Gabriele, La Malfa",
title="PreDA: Prefix-Based Dream Reports Annotation with Generative Language Models",
booktitle="Machine Learning, Optimization, and Data Science",
year="2026",
publisher="Springer Nature Switzerland",
address="Cham",
pages="191--206",
abstract="Dream reports are recollections of our experiences while asleep, and have strong research and clinical value. Since their analysis can be extremely time-consuming, researchers have adopted multiple types of automatised approaches, including, in more recent years, pre-trained language models (PLMs). However, most work has focused on limited aspects of the report content, such as characters or emotions. In this work, we introduce PreDA (prefix-based dream reports annotation), a framework to build language models able to annotate a dream report for multiple relevant aspects, using generative PLMs. We provide experimental evidence showing how a single PLM of small dimension can efficiently annotate a report on multiple features of the Hall and Van De Castle (HVDC) framework, give a detailed analysis of the model's performance, and explain how the training data impact learning and generalisation ability of the model.",
isbn="978-3-032-21477-5"
}