test_bastien.py

from datasets import load_dataset
from transformers import AutoTokenizer, AutoModel
from adapters import AutoAdapterModel, AdapterConfig
import re

def load_hate_speech_dataset():
    """
    Load the hate speech dataset from the UCBerkeley DLab.
    """
    mhs = load_dataset("ucberkeley-dlab/measuring-hate-speech", split='train')
    
    return mhs

def dataset_test():
    mhs = load_hate_speech_dataset()
    mhs = load_dataset("ucberkeley-dlab/measuring-hate-speech", split="train")

    # 2. Convert to a DataFrame for easy grouping
    df = mhs.to_pandas()

    # 3. Group by comment_id
    agg_df = (
        df.groupby("comment_id")
        .agg(
            hate_speech_score=("hate_speech_score", "mean"),  # average score
            platform=("platform", "first"),                   # same for every annotator row
            text=("text", "first")                            # same for every annotator row
        )
        .reset_index()
    )

    platform_names = {0: "Reddit", 1: "Reference", 2: "Twitter", 3: "Youtube"}
    platform_names = {0: "Reddit", 1: "Reference", 2: "Twitter", 3: "Youtube"}

    # Count comments per platform
    

    keyword = "rt"                     # ← put any word or phrase here

    pattern = rf"\b{re.escape(keyword)}\b"
    mask = agg_df["text"].str.contains(pattern, case=False, na=False, regex=True)
    hits = agg_df[mask]

    platform_counts = (
        agg_df["platform"]
        agg_df["platform"]
            .value_counts()          # counts for each integer code
            .sort_index()
    )

    #print(f"{len(hits)} comments contain “{keyword}” (case‑insensitive).")
    #print(f"{len(hits)} comments contain “{keyword}” (case‑insensitive).")
    print(platform_counts)

def test_model():          # adapter-aware class
    model = AutoAdapterModel.from_pretrained("roberta-base")

    # 2 – add a Houlsby bottleneck adapter
    cfg = AdapterConfig.load("houlsby", reduction_factor=16)
    model.add_adapter("hate_reg", config=cfg)

    # 3 – add a 768→1 linear head (treated as “classification” with num_labels=1)
    model.add_classification_head(
        "hate_reg",         # head name, same as adapter name = convenient
        num_labels=1,       # 👉 one scalar output
        layers=1,           # one linear layer is enough
        activation_function="identity",   # keep it linear for regression
    )

    # 4 – freeze backbone, keep only adapter + head trainable
    model.train_adapter(["hate_reg"])
    model.set_active_adapters("hate_reg")   
    
    print(model)  
    print("Active head:", model.active_head)   
    total_params     = sum(p.numel() for p in model.parameters())
    trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)

    print(f"{trainable_params:,} / {total_params:,} parameters are trainable "
        f"→ {trainable_params / total_params:.2%} of the model.")  

    print(model.heads["hate_reg"].config)  

        f"→ {trainable_params / total_params:.2%} of the model.")  

    print(model.heads["hate_reg"].config)  



if __name__ == "__main__":
    dataset_test()