Update README.md

README.md

@@ -20,4 +20,38 @@ widget:
 
 # Multilingual parliament sentiment regression model XLM-R-Parla-Sent
 
-This model is a result of a pre-training campaign, which trained XLM-Roberta-large on parliamentary transcriptions (collected in ParlaMint and EuroParl corpora), and fine-tuned on manually annotated sentiment datasets from United Kingdom, Czechia, Slovakia, Slovenia, Bosnia and Herzegovina, Croatia, and Serbia.
+This model is based on [xlm-r-parla](https://huggingface.co/classla/xlm-r-parla) and fine-tuned on manually annotated sentiment datasets from United Kingdom, Czechia, Slovakia, Slovenia, Bosnia and Herzegovina, Croatia, and Serbia.
+
+## Annotation schema
+
+The discrete labels, present in the original dataset, were mapped to integers as follows:
+
+```
+  "Negative": 0.0,
+  "M_Negative": 1.0,
+  "N_Neutral": 2.0,
+  "P_Neutral": 3.0,
+  "M_Positive": 4.0,
+  "Positive": 5.0,
+```
+Model was then fine-tuned on numeric labels and setup as regressor.
+
+
+
+## Finetuning procedure
+
+The fine-tuning procedure is described in this paper (ARXIV SUBMISSION to be added). Presumed optimal hyperparameters used are
+```
+  num_train_epochs=4,
+  train_batch_size=32,
+  learning_rate=8e-6,
+  regression=True
+```
+
+## Results
+
+Results reported were obtained from 10 fine-tuning runs.
+
+test dataset | R^2 
+BCS | 0.6146 ± 0.0104
+EN | 0.6722 ± 0.0100