eagle0504/context-engineering-v1

Context Engineering V1: Sequential API Recommendation Dataset

This dataset accompanies the research paper:

Rethink Context Engineering Using an Attention-based Architecture Yiqiao Yin — University of Chicago Booth School of Business / Columbia University

It was generated using the open-source context-engineer Python package:

• GitHub: https://github.com/yiqiao-yin/context-engineer-repo
• PyPI: https://pypi.org/project/context-engineer/0.1.0/

---

Dataset Summary

This dataset contains simulated sequential API usage logs modeled as Markov chains, designed for training and evaluating multi-task transformer models for sequential API recommendation. The simulation encompasses 2,000 user sessions totaling 20,000 API calls across 100 APIs organized into 10 functional categories, with 4 distinct session goal types driving workflow-specific behavioral patterns.

The dataset is split into two files:

File	Rows	Description
user_sessions.parquet	2,000	Full user session sequences with goal labels
training_pairs.parquet	18,000	Supervised input-output pairs for model training

Key Statistics

Metric	Value
Total users	2,000
Total API calls	20,000
Unique APIs	100 (across 10 categories)
Avg. session length	10 API calls
Session goal types	4
Training pairs generated	18,000
Max input sequence length	6
Random seed	42

---

Dataset Structure

user_sessions.parquet

Each row represents one complete user session:

Column	Type	Description
user_id	int	Unique user/session identifier (0–1999)
session_goal_id	int	Goal type ID (0–3)
session_goal	string	Goal name: ml_pipeline, data_analysis, user_management, quick_viz
sequence_length	int	Number of API calls in the session
api_sequence	string (JSON list)	Ordered list of API IDs called during the session
category_sequence	string (JSON list)	Ordered list of API category names

training_pairs.parquet

Each row is a supervised training example with multi-task labels:

Column	Type	Description
input_sequence	string (JSON list)	Context window of preceding API calls (up to 6)
input_length	int	Number of tokens in the input sequence
target_api	int	Ground-truth next API ID to predict
target_category	string	Category name of the target API
session_goal_id	int	Session goal label (auxiliary task)
session_goal	string	Session goal name
session_end	int	Whether this is the last action in the session (0 or 1)

---

API Categories

The 100 APIs are organized into 10 functional categories, reflecting typical enterprise platform architecture:

Category	API Range	Description
Authentication	0–9	Login, session management
User Management	10–19	Roles, permissions, accounts
Data Input	20–29	Data ingestion, file upload
Data Processing	30–39	Transformation, cleaning, feature engineering
ML Training	40–49	Model training, hyperparameter tuning
ML Prediction	50–59	Inference, batch prediction
Basic Visualization	60–69	Charts, basic plots
Advanced Visualization	70–79	Dashboards, interactive visualizations
Export/Share	80–89	Export, report generation
Administration	90–99	System config, monitoring

Session Goals

Goal ID	Goal Name	Distribution	Workflow Adherence
0	ML Pipeline	34.8%	85%
1	Data Analysis	26.1%	80%
2	User Management	24.3%	90%
3	Quick Visualization	14.8%	75%

---

How to Use

Load with Hugging Face datasets

from datasets import load_dataset

# Load both splits
dataset = load_dataset("eagle0504/context-engineering-v1")

# Or load individual files
sessions = load_dataset("eagle0504/context-engineering-v1", data_files="user_sessions.parquet")
pairs = load_dataset("eagle0504/context-engineering-v1", data_files="training_pairs.parquet")

Load with Pandas

import pandas as pd

sessions = pd.read_parquet("hf://datasets/eagle0504/context-engineering-v1/user_sessions.parquet")
pairs = pd.read_parquet("hf://datasets/eagle0504/context-engineering-v1/training_pairs.parquet")

Reproduce with the context-engineer Package

You can regenerate this exact dataset (or create your own variant) using the package:

pip install context-engineer

from context_engineer import simulate_multitask_markov_data, create_multitask_training_pairs, set_random_seeds

# Set seed for exact reproducibility
set_random_seeds(42)

# Generate 2000 user sessions (matches this dataset)
sequences, goals = simulate_multitask_markov_data(
    num_users=2000,
    num_apis=100,
    clicks_per_user=10,
)

# Create supervised training pairs
input_seqs, target_apis, goal_labels, session_end_labels = create_multitask_training_pairs(
    sequences, goals, max_seq_len=6
)

Run the Full Training Pipeline

from context_engineer import run_pipeline

# Reproduce the full experiment from the paper
results = run_pipeline(seed=42)

model = results["model"]       # Trained PyTorch model
metrics = results["metrics"]   # ~79.8% top-1 accuracy, 99.97% top-5 hit rate

Generate Custom Datasets via CLI

# Generate data and save to JSON
context-engineer generate --num-users 5000 --clicks 15 --seed 99 --output my_data.json

# Run the full pipeline
context-engineer run --num-users 1000 --epochs 30

---

Benchmark Results (from the paper)

A multi-task attention-based transformer trained on this dataset achieves:

Metric	Value
API Prediction Accuracy (Top-1)	79.83%
Mean Reciprocal Rank (MRR)	0.7983
Top-5 Hit Rate	99.97%
Top-10 Hit Rate	100.00%
Goal Prediction Accuracy	81.6%
Session End Accuracy	99.3%
Improvement over Markov baseline	+432%

---

Citation

If you use this dataset in your research, please cite:

@article{yin2025rethink,
  title={Rethink Context Engineering Using an Attention-based Architecture},
  author={Yin, Yiqiao},
  year={2025}
}

---

Disclaimer

About the Author. This dataset and the accompanying context-engineer package were created by Yiqiao Yin, who holds affiliations with the University of Chicago Booth School of Business and the Department of Statistics at Columbia University. The author brings over a decade of professional experience in the SaaS (Software as a Service) and Platform-as-a-Service (PaaS) domain, spanning enterprise software development, API ecosystem design, user behavior analytics, and machine learning infrastructure. The API category taxonomy, workflow patterns, user persona definitions, and transition probability structures encoded in this simulator are informed by that cumulative domain expertise—reflecting realistic patterns observed in production enterprise environments over the course of many years.

Simulation, Not Real Data. This dataset is entirely synthetic. It was generated programmatically using the open-source context-engineer Python package. No real user data, proprietary platform logs, personally identifiable information (PII), or third-party datasets of any kind are included, referenced, or derived from in this release. The Markov chain transition probabilities, user personas, and session goal distributions are designed to approximate realistic enterprise API usage patterns for research purposes, but they do not represent, reproduce, or leak any actual user behavior from any specific platform or organization.

Reproducibility. This dataset is fully reproducible. Running the generation script with seed=42 and the default parameters (num_users=2000, num_apis=100, clicks_per_user=10) will produce an identical dataset. The source code is publicly available at github.com/yiqiao-yin/context-engineer-repo.

License. This dataset is released under the MIT License. You are free to use, modify, and distribute it for academic and commercial purposes with attribution.