Update README.md

README.md

@@ -1 +1,3 @@
-models/roberta-base-ca-v2/wiki_no_teca/output/roberta-base-ca-v2/tecla_nli.py_8_0.00003_date_22-12-21_time_01-40-19/checkpoint-2494
+models/roberta-base-ca-v2/wiki_no_teca/output/roberta-base-ca-v2/tecla_nli.py_8_0.00003_date_22-12-21_time_01-40-19/checkpoint-2494
+
+In order to train the two entailment models, RoBERTa- ca-Wikicorpus and RoBERTa-ca-Teca-Wikicorpus, we first transformed the Wikicorpus dataset into the entailment format. In this process, to generate the hypothesis, we employed the same approach outlined in Section 3.1, utilizing the seventh template from Table 6 (“Aquest article tracta sobre {label}.”), which yielded consistently strong results across both the coarse-grained and the fine-grained tasks for our model. To balance the proportion of entailment and non-entailment hypotheses and avoid the computational cost of multiplying the dataset size by 67 classes (which would be the case if we generated all possible non- entailment hypotheses for each entailment hypothesis), we decided to generate one non- entailment per each entailment hypothesis, thereby only increasing the dataset size by a factor of two. For the training of the entailment models with Wikicorpus, we kept the configurations used in the main few-shot experiments previously presented in this section: we selected the best checkpoint according to the weighted F1 score in the classification task and kept the same fixed hyperparameters.