CT-SCaMS

Continous Time-Series Clustering and Merging towards Sentiments (CT-SCaMS) is a Python package ddesigned to ensure robust generalization across time stamps for sentiment analysis in policy-related texts.

It accounts for the temporal dynamics of evolving policy discourse and evaluates sentiment models under realistic settings that mimic actual policy analysis scenarios.

Key features include:

• Continuous time-series clustering to prioritize data points for annotation based on temporal trends.

• Advanced model merging techniques to unify multiple models fine-tuned on distinct temporal slices.

Methods:

• Continous time series is performed using Ruptures

• It supports the following merging techniques:

  1. Uniform Souping

  2. Greedy Souping

  3. Task Arithmetic

  4. TIES

  5. DARE

  6. Fisher Merging

  7. RegMean Merging

Requirements

Required Packages

The following Python packages are required:

• matplotlib

• ruptures

• pandas

• tqdm

• torch

• transformers

• datasets

• scikit-learn

• numpy

These will be installed automatically when using pip install.

Alternatively, you could directly install them

pip install matplotlib ruptures pandas tqdm torch transformers datasets scikit-learn numpy

Computational requirements

You require at least one GPU to use ctscams.

VRAM requirements depend on factors like batch size, model size, etc.

However, at least 12GB of VRAM is recommended

Installation

To install in python, simply do the following:

pip install ctscams

Quick start

To learn how we used it in our experiments, go to demo.ipynb. Alternatively, refer to this quick guide here:

1. Lets start off by performing continous time series clustering to select our annotated data

from ctscams import greedy_souping, ties

from ctscams import finetune

from ctscams import cluster_sampling, continous_time_series_clustering



df_climate=continous_time_series_clustering(df=df,time_col="timestamp_col",level="M", plot=False, penalty=0.1) # note that level "M" stands for Month



# perform continous time series clustering and selecting based on the clusters

df=cluster_sampling(df=df,sample_size=10000,stratified_col="cluster"):

df_annotated=df[df["selected"]==1].reset_index(drop=True)

2. Now first lets finetune a single unified model

# finetuning

finetune(df=df_annotated, model_name="microsoft/deberta-v3-large",cluster_col_name=None,

         folder_name="naive_finetuning/continous_clustering_by_month/climate_change", # note folder name to save our models

         text_col='message', label_col="sentiment",  label2id={"Anti":0,"Neutral":1,"Pro":2,"News":3},

         learning_rate=1e-5, warmup_ratio=0.05, weight_decay=0.001,

         epochs=3, batch_size=6, early_stopping_patience=2, return_val_data=False)

3. Now first lets finetune individual models (one for each cluster), and merge them with greedy soup.

# finetuning

df_val=finetune(df=df_annotated, model_name="microsoft/deberta-v3-large",

                cluster_col_name="cluster", # the difference is here --> we use "cluster" instead of None

                folder_name="batch_finetuning/continous_clustering_by_month/climate_change", # note folder name to save our models

                text_col='message', 

                label_col="sentiment",  

                label2id={"Anti":0,"Neutral":1,"Pro":2,"News":3},

                learning_rate=1e-5, 

                warmup_ratio=0.05, 

                weight_decay=0.001,

                epochs=8, 

                batch_size=6, 

                early_stopping_patience=2, 

                return_val_data=True) # note TRUE if we want to return validation data. 



# Now we perform greedy souping

greedy_souping(

    models_folder="models/batch_finetuning/continous_clustering_by_month/climate_change/deberta-v3-large", # where the individual models are saved

    save_path="models/merged_models/continous_clustering_by_month/climate_change/greedy_soup/deberta-v3-large", # the merged model will be saved here

    df_val=df_val,  # validation data. 

    col_label="sentiment",

    text_col='message',

    num_labels=4,

    label2id={"Anti":0,"Neutral":1,"Pro":2,"News":3})



# the model will be saved under `save_path="models/merged_models/continous_clustering_by_month/climate_change/greedy_soup/deberta-v3-large"`

Reference and citation

Alba C, Warner BC, Saxena A, Huang J, An R. Towards Robust Sentiment Analysis of Temporally-Sensitive Policy-Related Online Text. Accepted In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL), Volume 4: Student Research Workshop. Association for Computational Linguistics; July 26-August 1, 2025; Vienna, Austria.