arafatanam/Student-Mental-Health-Counseling-EN

Student Mental Health Counseling (EN)

Dataset Overview

This dataset is the final cleaned and filtered version in a three-step pipeline that started with the original Vietnamese counseling dataset.

Step	Dataset	Rows
1. Original	chillies/student-mental-health-counseling-vn	750,169
2. Translated	arafatanam/Student-Mental-Health-Counseling-750K	750,169
3. Filtered (this dataset)	arafatanam/Student-Mental-Health-Counseling-EN	52,254

The goal of this version is to keep only conversations that are directly relevant to student mental health, removing general counseling data that does not relate to academic life, school, or student experiences.

Dataset Details

• Dataset Name: Student Mental Health Counseling (EN)
• Source Dataset: arafatanam/Student-Mental-Health-Counseling-750K
• Language: English
• Final Row Count: 52,254
• Format: CSV
• Columns: Question, Answer

Processing Pipeline

Step 1: Translation (done in parent dataset)

The original Vietnamese dataset was fully translated to English using Helsinki-NLP/opus-mt-vi-en (MarianMT) with checkpoint-based translation to handle the full 750,169 rows without data loss.

Step 2: Filtering (done in this dataset)

The filtering runs in three stages:

Stage 1: Null and empty row removal Rows with missing or whitespace-only values in either Question or Answer are dropped immediately.

Stage 2: Exact duplicate removal Rows where both Question and Answer are identical are removed before embedding to avoid wasting compute. This removed 401,812 rows (53% of the translated dataset), indicating heavy repetition in the original scraped source.

Stage 3: Semantic similarity filtering Each row is embedded using sentence-transformers/all-MiniLM-L6-v2. The cosine similarity is computed against the mean embedding of 20 curated seed sentences representing student mental health conversations. The top 15% of rows by similarity score are kept (cutoff: 0.2278).

This approach captures meaning rather than exact keywords, so conversations like "I freeze during assessments" are retained even without explicit academic vocabulary.

Filtering Code

# Install ──────────────────────────────────────────
# !pip install datasets huggingface_hub sentence-transformers pandas -q

# Imports + Auth ───────────────────────────────────
import os
import torch
import numpy as np
import pandas as pd
from datasets import load_dataset
from huggingface_hub import login
from sentence_transformers import SentenceTransformer
from kaggle_secrets import UserSecretsClient

secrets = UserSecretsClient()
HF_TOKEN = secrets.get_secret("HF_TOKEN")
login(token=HF_TOKEN)

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
GPU_COUNT = torch.cuda.device_count()
print(f"Device      : {DEVICE}")
print(f"GPU count   : {GPU_COUNT}")
if DEVICE == "cuda":
    for i in range(GPU_COUNT):
        print(f"  GPU {i}    : {torch.cuda.get_device_name(i)}")

# Load Dataset ─────────────────────────────────────
print("\nLoading dataset...")
dataset = load_dataset(
    "arafatanam/Student-Mental-Health-Counseling-750K",
    token=HF_TOKEN,
)
df = dataset["train"].to_pandas()

df = df.dropna(subset=["Question", "Answer"]).reset_index(drop=True)
df["Question"] = df["Question"].str.strip()
df["Answer"]   = df["Answer"].str.strip()
df = df[(df["Question"] != "") & (df["Answer"] != "")].reset_index(drop=True)
print(f"Rows after null/empty drop : {len(df):,}")

# Remove Exact Duplicates Early ────────────────────
before_dedup = len(df)
df = df.drop_duplicates(subset=["Question", "Answer"]).reset_index(drop=True)
after_dedup = len(df)
print(f"Exact duplicates removed   : {before_dedup - after_dedup:,}")
print(f"Rows remaining             : {after_dedup:,}")

# Load Embedding Model ─────────────────────────────
print("\nLoading embedding model...")
model = SentenceTransformer(
    "sentence-transformers/all-MiniLM-L6-v2",
    device=DEVICE,
)
print("Model loaded.")

# Define Seed Sentences ────────────────────────────
SEED_SENTENCES = [
    # Direct student experiences
    "I am a university student feeling overwhelmed by exam pressure",
    "I cannot sleep because I am anxious about my upcoming exams",
    "My grades are dropping and I feel like a failure",
    "I feel so much pressure to perform well academically",
    "I am a college student struggling with depression and loneliness",
    "Balancing coursework and mental health is becoming impossible",
    "I am scared of failing my classes this semester",
    "My academic performance is suffering because of my anxiety",
    "I feel burned out from studying all the time",
    "I do not know how to cope with the stress of school",
    # Academic environment issues
    "I have a difficult relationship with my professor",
    "My classmates bully me and I feel excluded on campus",
    "I am struggling to keep up with my assignments and deadlines",
    "I feel pressure from my parents to get good grades",
    "I am considering dropping out because I cannot handle the stress",
    "I moved into the dormitory and feel extremely homesick",
    "I have test anxiety and I freeze during exams",
    "I do not understand the coursework and I am too afraid to ask for help",
    "I feel like I am the only student who is struggling",
    "Scholarship pressure is affecting my mental health severely",
]

print(f"\nEncoding {len(SEED_SENTENCES)} seed sentences...")
seed_embeddings = model.encode(
    SEED_SENTENCES,
    batch_size=64,
    show_progress_bar=False,
    convert_to_numpy=True,
    normalize_embeddings=True,
)
seed_mean = seed_embeddings.mean(axis=0, keepdims=True)
print("Seed embedding computed.")

# Embed Dataset Rows ───────────────────────────────
BATCH_SIZE = 512
texts = (df["Question"] + " [SEP] " + df["Answer"]).tolist()

print(f"\nEmbedding {len(texts):,} rows in batches of {BATCH_SIZE}...")
row_embeddings = model.encode(
    texts,
    batch_size=BATCH_SIZE,
    show_progress_bar=True,
    device=DEVICE,
    convert_to_numpy=True,
    normalize_embeddings=True,
)
print(f"\nEmbedding complete. Shape: {row_embeddings.shape}")

# Compute Similarities ─────────────────────────────
similarities = (row_embeddings @ seed_mean.T).squeeze()
df["similarity_score"] = similarities

# Threshold Analysis ───────────────────────────────
print(f"{'Threshold':>12} | {'Rows Kept':>12} | {'% of Total':>12}")
print("-" * 42)
for t in np.arange(0.20, 0.55, 0.05):
    count = int((similarities >= t).sum())
    pct   = count / len(df) * 100
    print(f"   {t:.2f}      |  {count:>10,}  |  {pct:>9.1f}%")

# Percentile-Based Filter ─────────────────────────
TOP_PERCENT = 15
cutoff      = np.percentile(similarities, 100 - TOP_PERCENT)
mask        = similarities >= cutoff
df_filtered = df[mask].drop(columns=["similarity_score"]).reset_index(drop=True)

print(f"  Method         : Top {TOP_PERCENT}% by similarity")
print(f"  Score cutoff   : {cutoff:.4f}")
print(f"  Filtered rows  : {len(df_filtered):,}")

# Save to Kaggle Output ───────────────────────────
OUTPUT_DIR = "/kaggle/working"

parquet_path = f"{OUTPUT_DIR}/student_mh.parquet"
df_filtered.to_parquet(parquet_path, index=False)

csv_path = f"{OUTPUT_DIR}/student_mh.csv"
df_filtered.to_csv(csv_path, index=False)

print(f"Saved parquet : {parquet_path}")
print(f"Saved CSV     : {csv_path}")
print(f"Total rows    : {len(df_filtered):,}")

Pipeline Statistics

Stage	Rows	Change
Translated dataset loaded	750,169
After null/empty drop	750,169	0 removed
After exact dedup	348,357	401,812 removed
After semantic filter (top 15%)	52,254	296,103 removed

Similarity Score Distribution

Computed against 20 student mental health seed sentences using all-MiniLM-L6-v2 on a T4 GPU (Kaggle).

Threshold	Rows Kept	% of Total
0.20	91,302	26.2%
0.25	30,906	8.9%
0.30	6,594	1.9%
0.35	859	0.2%

A percentile-based cutoff (top 15%, score >= 0.2278) was chosen over a fixed threshold to account for translation-induced score compression in the embedded text.

Intended Use

• Fine-tuning LLMs for student-focused mental health support chatbots
• Research on academic stress, burnout, and student wellbeing
• Benchmarking counseling response quality in academic contexts

License

Apache 2.0, inherited from the original dataset.

Citation

Please credit all datasets in the pipeline:

1. chillies/student-mental-health-counseling-vn (original Vietnamese source)
2. arafatanam/Student-Mental-Health-Counseling-750K (full English translation)
3. arafatanam/Student-Mental-Health-Counseling-EN (this filtered dataset)