autom4ta/cd-erc-roberta-dailydialog

Checkpoint do cd-erc (Conversational emotion Detection, roberta-base

• contexto de turnos, DailyDialog) treinado no cluster do CEIA (execucao-ceia/, job 29802, epoca 4). Baseado em HLT-MAIA/Emotion-Transformer.

Nao carrega com AutoModelForSequenceClassification.from_pretrained. O modelo e um pl.LightningModule customizado (encoder roberta-base com embeddings redimensionados para 3 tokens especiais <bos>/<eos>/<pad>, cabeca nn.Linear propria, forward(input_ids, input_lengths) nao-padrao) — nao e um *ForSequenceClassification padrao do transformers.

Como carregar

Via o servico do cd-erc-module (recomendado):

CDERC_HF_REPO=autom4ta/cd-erc-roberta-dailydialog ./run.sh up cderc

ou direto no codigo do modulo (cd-erc-module/):

from pathlib import Path
from huggingface_hub import hf_hub_download
from model.emotion_transformer import EmotionTransformer

ckpt = hf_hub_download(repo_id="autom4ta/cd-erc-roberta-dailydialog", filename="checkpoints/model.ckpt")
hf_hub_download(repo_id="autom4ta/cd-erc-roberta-dailydialog", filename="hparams.yaml")
folder = str(Path(ckpt).parents[1]) + "/"
model = EmotionTransformer.from_experiment(folder)

Metricas (execucao-ceia, split de teste do DailyDialog)

metrica	valor
macro-f1	0,5163
accuracy	0,8541 (enganosa — no emotion e 81,7% das amostras)

Config de treino (hparams.yaml)

• pretrained_model: roberta-base
• labels: dailydialog (7 classes: no emotion, anger, disgust, fear, happiness, sadness, surprise)
• context: true, context_turns: 3