Hugging Face's logo

Datasets:
Omarrran
/
stackpulse_qa_output

Tasks:
Question Answering
Text Generation
Modalities:
Tabular
Text
Languages:
English
Size:
100K - 1M
Tags:
stackoverflow
instruction-tuning
qa
code
fine-tuning
alpaca-format
License:
Dataset card Data Studio Files
xet
 Community
stackpulse_qa_output / README.md
Omarrran's picture
Omarrran
Upload README.md
b5826d2 verified
preview code
|
raw
history blame contribute delete
6.5 kB
metadata 
license: apache-2.0
language:
  - en
pretty_name: 'StackPulse-QA: Instruction-Tuning Q&A Pairs from Stack Overflow'
size_categories:
  - 100K<n<1M
task_categories:
  - question-answering
  - text-generation
  - text2text-generation
tags:
  - stackoverflow
  - instruction-tuning
  - qa
  - code
  - fine-tuning
  - alpaca-format
  - llm-training

🧩 StackPulse-QA: Instruction-Tuning Q&A Pairs from Stack Overflow

Dataset Summary

Instruction-tuning Q&A dataset built from Omarrran/StackPulse_778K_QnA_Code_dataset by joining question IDs with BigQuery bigquery-public-data.stackoverflow.posts_answers on accepted_answer_id.

Each sample consists of:

  • input_text_instruct — A question (title + body) prefixed with an instruction
  • output_text — The accepted answer from Stack Overflow

Format mirrors the instruction-tuning dataset from DeepLearning.AI's Finetuning Large Language Models course, ready for fine-tuning PaLM, LLaMA, Mistral, Gemma, Phi, and similar models.

📊 Processing Progress

  • Runs completed : 4 / 6
  • Questions processed : 400,000 / 554,196
  • Remaining : 154,196

📁 Files in This Dataset

🏋️ Training Files (80% split)

File Format Description
data/tune_data_stack_overflow_python_qa_run1-07:19:04:2026.jsonl JSONL Training split from 1
data/tune_data_stack_overflow_python_qa_run2-07:19:04:2026.jsonl JSONL Training split from 2
data/tune_data_stack_overflow_python_qa_run3-07:19:04:2026.jsonl JSONL Training split from 3
data/tune_data_stack_overflow_python_qa_run4-07:19:04:2026.jsonl JSONL Training split from 4
data/tune_data_stack_overflow_python_qa_run5-07:19:04:2026.jsonl JSONL Training split from 5

🧪 Evaluation Files (20% split)

File Format Description
data/tune_eval_data_stack_overflow_python_qa_run1-07:19:04:2026.jsonl JSONL Eval split from run 1
data/tune_eval_data_stack_overflow_python_qa_run2-07:19:04:2026.jsonl JSONL Eval split from run 2
data/tune_eval_data_stack_overflow_python_qa_run3-07:19:04:2026.jsonl JSONL Eval split from run 3
data/tune_eval_data_stack_overflow_python_qa_run4-07:19:04:2026.jsonl JSONL Eval split from run 4

📄 Full Metadata CSVs

File Format Description
data/stackpulse_qa_full_run1-07:19:04:2026.csv CSV Full metadata for run 1
data/stackpulse_qa_full_run2-07:19:04:2026.csv CSV Full metadata for run 2
data/stackpulse_qa_full_run3-07:19:04:2026.csv CSV Full metadata for run 3
data/stackpulse_qa_full_run4-07:19:04:2026.csv CSV Full metadata for run 4

🏗️ Schema

JSONL Files (training / eval)

Exactly 2 fields per row — ready for instruction fine-tuning:

Field Type Description
input_text_instruct string Instruction prefix + question title + question body
output_text string Accepted answer body (HTML format)

CSV Files (full metadata)

Column Description
question_id Stack Overflow question ID
input_text title + body (no instruction prefix)
output_text accepted answer body
input_text_instruct instruction-prefixed input (same as JSONL)
title question title only
tags pipe-separated tags
q_score question upvote score
view_count total views
answer_count number of answers
accepted_answer_id ID of the accepted answer
answer_id ID of this answer (= accepted_answer_id)
a_score answer upvote score
is_accepted always True (we only keep accepted answers)
creation_date question creation timestamp

🚀 Quick Start

Load with pandas 

import pandas as pd

# Training data
train = pd.read_json("data/tune_data_stack_overflow_python_qa_run1-*.jsonl", lines=True)

# Eval data
eval_  = pd.read_json("data/tune_eval_data_stack_overflow_python_qa_run1-*.jsonl", lines=True)

print(train.iloc[0]["input_text_instruct"][:300])
print(train.iloc[0]["output_text"][:300])

Load with HuggingFace datasets 

from datasets import load_dataset

# Load all training shards
ds = load_dataset(
    "json",
    data_files={
        "train": "data/tune_data_stack_overflow_python_qa_run*.jsonl",
        "eval"  : "data/tune_eval_data_stack_overflow_python_qa_run*.jsonl",
    }
)
print(ds)

Use for fine-tuning (Alpaca-style) 

def format_prompt(ex):
    return {
        "text": f"{ex['input_text_instruct']}\n\n### Response:\n{ex['output_text']}"
    }

train_formatted = ds["train"].map(format_prompt)

📋 Instruction Template Used

Please answer the following Stackoverflow question on Programming. Answer it like you are a developer answering Stackoverflow questions. Stackoverflow question: {title}{body}

⚠️ Caveats

  1. HTML in answers: output_text contains raw HTML tags (<p>, <pre>, <code>). Strip or preserve depending on your use case.
  2. Accepted answers only: We filter q.accepted_answer_id = a.id — other community answers are skipped.
  3. ~60% match rate: Of each 100K question IDs queried, ~60K have accepted answers in BigQuery. The rest are self-answered, deleted, or lack acceptance.
  4. 80/20 split: Each run uses random_state=42 for reproducible train/eval splits.
  5. Mirrors L2_data.ipynb: Format exactly matches DeepLearning.AI's Finetuning Large Language Models course notebook structure.

🔁 Source Dataset

Question IDs and metadata sourced from:

Answers joined from:

  • bigquery-public-data.stackoverflow.posts_answers (Google BigQuery Public Dataset)

📋 Citation 

@dataset{malik2026stackpulseqa,
  author    = {Malik, Omar Haq Nawaz},
  title     = {StackPulse-QA: Instruction-Tuning Q&A Pairs from Stack Overflow},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/Omarrran/stackpulse_qa_output},
  license   = {Apache-2.0}
}

👤 Author

Omar Haq Nawaz Malik (HuggingFace: Omarrran) AI Engineer & NLP Researcher | BITS Pilani | Srinagar, Kashmir