Upload README.md with huggingface_hub

README.md

@@ -1,3 +1,499 @@
----
-license: mit
----
+---
+license: cc-by-4.0
+task_categories:
+- other
+tags:
+- physics
+- high-energy-physics
+- particle-physics
+- tracking
+- calorimetry
+- machine-learning
+- simulation
+pretty_name: ColliderML Top-Quark Pair Production (No Pileup)
+size_categories:
+- 10K<n<100K
+configs:
+- config_name: particles
+  data_files: "https://portal.nersc.gov/cfs/m4958/ColliderML/hard_scatter/ttbar/v1/parquet/truth/particles/*.parquet"
+- config_name: tracker_hits
+  data_files: "https://portal.nersc.gov/cfs/m4958/ColliderML/hard_scatter/ttbar/v1/parquet/reco/tracker_hits/*.parquet"
+- config_name: calo_hits
+  data_files: "https://portal.nersc.gov/cfs/m4958/ColliderML/hard_scatter/ttbar/v1/parquet/reco/calo_hits/*.parquet"
+- config_name: tracks
+  data_files: "https://portal.nersc.gov/cfs/m4958/ColliderML/hard_scatter/ttbar/v1/parquet/reco/tracks/*.parquet"
+---
+
+# ColliderML: Top-Quark Pair Production Dataset (ttbar, No Pileup)
+
+## Dataset Description
+
+This dataset contains simulated high-energy physics collision events for top-quark pair (ttbar) production with **no pileup** (single interaction per event). The data is generated using the **Open Data Detector (ODD)** geometry within the **ACTS (A Common Tracking Software)** framework, representing a generic collider detector similar to those at the LHC.
+
+### Dataset Summary
+
+- **Campaign**: `hard_scatter`
+- **Process**: Top-quark pair production (ttbar)
+- **Version**: `v1`
+- **Number of Events**: ~29,000 events (29 files × 1000 events per file)
+- **Pileup**: 0 (no additional interactions)
+- **Detector**: Open Data Detector (ODD)
+- **Format**: Apache Parquet with list columns for variable-length data
+- **License**: CC-BY-4.0
+
+### Supported Tasks
+
+This dataset is designed for machine learning tasks in high-energy physics, including:
+
+- **Particle tracking**: Reconstruct charged particle trajectories from detector hits
+- **Track-to-particle matching**: Associate reconstructed tracks with truth particles
+- **Jet tagging**: Identify jets originating from top quarks, b-quarks, or light quarks
+- **Energy reconstruction**: Predict particle energies from calorimeter deposits
+- **Physics analysis**: Event classification (signal vs. background discrimination)
+- **Representation learning**: Study hierarchical information at different detector levels
+
+### Languages
+
+N/A (Physics data)
+
+## Dataset Structure
+
+### Data Instances
+
+Each row in the Parquet files represents a single collision event. Variable-length quantities (e.g., lists of particles, hits, tracks) are stored as Parquet list columns.
+
+Example event structure:
+
+```python
+{
+    'event_id': 42,
+    'particle_id': [0, 1, 2, 3, ...],  # List of particle IDs
+    'pdg_id': [11, -11, 211, ...],     # Particle type codes
+    'px': [1.2, -0.5, 3.4, ...],       # Momentum components (GeV)
+    'py': [0.8, 1.1, -0.3, ...],
+    'pz': [5.2, -2.1, 10.5, ...],
+    'energy': [5.5, 2.3, 11.2, ...],
+    # ... additional fields
+}
+```
+
+### Data Fields
+
+The dataset contains four data types organized by detector hierarchy:
+
+#### 1. `particles` (Truth-level)
+
+Truth information about generated particles before detector simulation.
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `event_id` | int64 | Unique event identifier |
+| `particle_id` | list\<int64\> | Unique particle ID within event |
+| `pdg_id` | list\<int64\> | PDG particle code (e.g., 11=electron, 13=muon, 211=pion) |
+| `mass` | list\<float64\> | Particle rest mass (GeV/c²) |
+| `energy` | list\<float64\> | Particle total energy (GeV) |
+| `charge` | list\<float64\> | Electric charge (in units of e) |
+| `px`, `py`, `pz` | list\<float64\> | Momentum components (GeV/c) |
+| `vx`, `vy`, `vz` | list\<float64\> | Vertex position (mm) |
+| `time` | list\<float64\> | Production time (ns) |
+| `num_tracker_hits` | list\<int64\> | Number of hits in tracker |
+| `num_calo_hits` | list\<int64\> | Number of hits in calorimeter |
+| `vertex_primary` | list\<int64\> | Primary vertex flag (1=primary, 0=secondary) |
+| `parent_id` | list\<float64\> | ID of parent particle |
+
+**Typical event**: ~200-300 particles per event
+
+#### 2. `tracker_hits` (Detector-level)
+
+Digitized spatial measurements from the tracking detector (silicon sensors).
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `event_id` | int64 | Unique event identifier |
+| `x`, `y`, `z` | list\<float64\> | Measured hit position (mm) |
+| `true_x`, `true_y`, `true_z` | list\<float64\> | True (simulated) hit position before digitization (mm) |
+| `time` | list\<float64\> | Hit time (ns) |
+| `particle_id` | list\<int64\> | Truth particle that created this hit |
+| `volume_id` | list\<int64\> | Detector volume identifier |
+| `layer_id` | list\<int64\> | Detector layer number |
+| `surface_id` | list\<int64\> | Sensor surface identifier |
+| `cell_id` | list\<int64\> | Cell/pixel identifier |
+| `detector` | list\<int64\> | Detector subsystem code |
+
+**Typical event**: ~2,000-3,000 hits per event
+
+#### 3. `calo_hits` (Calorimeter-level)
+
+Energy deposits in the calorimeter system (electromagnetic + hadronic).
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `event_id` | int64 | Unique event identifier |
+| `detector` | list\<string\> | Calorimeter subsystem name |
+| `cell_id` | list\<string\> | Calorimeter cell identifier |
+| `total_energy` | list\<float64\> | Total energy deposited in cell (GeV) |
+| `x`, `y`, `z` | list\<float64\> | Cell center position (mm) |
+| `contrib_particle_ids` | list\<list\<int64\>\> | IDs of particles contributing to this cell |
+| `contrib_energies` | list\<list\<float64\>\> | Energy contribution from each particle (GeV) |
+| `contrib_times` | list\<list\<float64\>\> | Time of each contribution (ns) |
+
+**Note**: Nested lists for contributions (one cell can have multiple particle deposits).
+
+**Typical event**: ~500-1,000 calorimeter cells with deposits
+
+#### 4. `tracks` (Reconstruction-level)
+
+Reconstructed particle tracks from pattern recognition and track fitting algorithms.
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `event_id` | int64 | Unique event identifier |
+| `track_id` | list\<int64\> | Unique track identifier within event |
+| `majority_particle_id` | list\<int64\> | Truth particle with most hits on this track |
+| `d0` | list\<float64\> | Transverse impact parameter (mm) |
+| `z0` | list\<float64\> | Longitudinal impact parameter (mm) |
+| `phi` | list\<float64\> | Azimuthal angle (radians) |
+| `theta` | list\<float64\> | Polar angle (radians) |
+| `qop` | list\<float64\> | Charge divided by momentum (e/GeV) |
+| `hit_ids` | list\<list\<int32\>\> | List of tracker hit IDs assigned to this track |
+
+**Track parameters**: Standard ACTS track representation (perigee parameters at origin).
+
+**Derived quantities**:
+- Transverse momentum: `pt = abs(1/qop) * sin(theta)`
+- Pseudorapidity: `eta = -ln(tan(theta/2))`
+- Total momentum: `p = abs(1/qop)`
+
+**Typical event**: ~100-150 reconstructed tracks per event
+
+### Data Splits
+
+Currently, the dataset does not have predefined train/validation/test splits. Users should implement their own splitting strategy based on their use case. Recommended approach:
+
+```python
+from sklearn.model_selection import train_test_split
+
+# Example: 70% train, 15% validation, 15% test
+all_events = list(range(29000))
+train_val, test = train_test_split(all_events, test_size=0.15, random_state=42)
+train, val = train_test_split(train_val, test_size=0.176, random_state=42)  # 0.176 * 0.85 ≈ 0.15
+```
+
+## Dataset Creation
+
+### Curation Rationale
+
+This dataset was created to support machine learning research in high-energy physics, specifically for:
+
+1. **Benchmarking tracking algorithms**: Compare traditional and ML-based track reconstruction methods
+2. **Hierarchical representation learning**: Study information flow from detector hits → tracks → particles
+3. **Physics analysis**: Develop ML models for event classification and particle identification
+4. **Open science**: Provide publicly accessible, realistic detector simulation data
+
+The ttbar process is chosen because:
+- It produces complex final states with many particles
+- It's a key signature at hadron colliders (LHC)
+- Top quarks decay to b-quarks, W bosons, and ultimately jets and leptons
+- Relevant for searches for new physics beyond the Standard Model
+
+### Source Data
+
+#### Initial Data Collection and Normalization
+
+The data is generated through the following simulation chain:
+
+1. **Event Generation**: ttbar events generated using a Monte Carlo event generator
+2. **Detector Simulation**: Particle propagation through the Open Data Detector using ACTS
+3. **Digitization**: Conversion of energy deposits to realistic detector signals
+4. **Reconstruction**: Track finding and fitting using ACTS tracking algorithms
+5. **Format Conversion**: EDM4HEP → Parquet using the ColliderML data pipeline
+
+#### Who are the source data producers?
+
+The data is produced by the **ColliderML collaboration** as part of the **ATLAS ITk ML Reconstruction** project at NERSC (National Energy Research Scientific Computing Center).
+
+### Annotations
+
+#### Annotation process
+
+The dataset includes truth-level annotations automatically generated during the simulation:
+
+- **Particle-level truth**: Generator-level particle information
+- **Hit-to-particle associations**: Which particle created each detector hit
+- **Track-to-particle matching**: `majority_particle_id` links reconstructed tracks to truth particles
+
+These annotations enable supervised learning for tasks like:
+- Track efficiency (did we reconstruct this particle?)
+- Track purity (how many hits belong to the correct particle?)
+- Fake rate (how many tracks are not matched to real particles?)
+
+#### Who are the annotators?
+
+N/A (Annotations are from simulation ground truth)
+
+### Personal and Sensitive Information
+
+This dataset contains only simulated physics data. No personal or sensitive information is included.
+
+## Considerations for Using the Data
+
+### Social Impact of Dataset
+
+This dataset supports fundamental physics research and ML algorithm development. It has no direct social impact but contributes to:
+
+- Open science and reproducible research
+- Education in HEP and ML
+- Development of algorithms that may have broader applications (e.g., pattern recognition, tracking in medical imaging)
+
+### Discussion of Biases
+
+As a simulated dataset, biases may arise from:
+
+1. **Generator-level biases**: The event generator's modeling of ttbar production
+2. **Detector simulation biases**: Approximations in material interactions, detector response
+3. **Reconstruction biases**: Algorithm choices in track finding and fitting
+4. **No pileup**: Real LHC data has 20-60 simultaneous collisions; this dataset has only 1
+
+Users should be aware that models trained on this data may not generalize to:
+- Real detector data (requires calibration and alignment)
+- Different detector geometries
+- Events with pileup
+
+### Other Known Limitations
+
+- **Limited statistics**: ~29,000 events is moderate for ML training (consider data augmentation)
+- **Single physics process**: Only ttbar; does not include background processes
+- **Idealized detector**: ODD is a generic detector, not an exact replica of ATLAS/CMS
+- **No detector inefficiencies**: Assumes 100% hit efficiency (real detectors have dead regions)
+
+## Additional Information
+
+### Dataset Curators
+
+This dataset is maintained by the ColliderML team:
+
+- Primary contact: [danieltm@lbl.gov](mailto:danieltm@lbl.gov)
+- Collaboration: ATLAS ITk ML Reconstruction working group
+- Infrastructure: NERSC (National Energy Research Scientific Computing Center)
+
+### Licensing Information
+
+This dataset is released under the **Creative Commons Attribution 4.0 International (CC BY 4.0)** license.
+
+You are free to:
+- **Share**: Copy and redistribute the material
+- **Adapt**: Remix, transform, and build upon the material
+
+Under the following terms:
+- **Attribution**: You must give appropriate credit and indicate if changes were made
+
+### Citation Information
+
+If you use this dataset in your research, please cite:
+
+```bibtex
+@dataset{colliderml_ttbar_pu0_2024,
+  title={ColliderML: Top-Quark Pair Production Dataset (No Pileup)},
+  author={ColliderML Collaboration},
+  year={2024},
+  publisher={NERSC},
+  howpublished={\url{https://huggingface.co/datasets/OpenDataDetector/ColliderML_ttbar_pu0}},
+  note={Simulation performed using ACTS and the Open Data Detector}
+}
+```
+
+### Contributions
+
+This dataset was produced using:
+
+- **ACTS (A Common Tracking Software)**: https://acts.readthedocs.io/
+- **Open Data Detector**: https://acts.readthedocs.io/en/latest/examples/open_data_detector.html
+- **EDM4HEP**: https://edm4hep.web.cern.ch/
+- **ColliderML Pipeline**: https://github.com/ATLAS-ITk-ML/colliderml
+
+## How to Use This Dataset
+
+### Loading the Dataset
+
+The dataset is hosted on the NERSC public portal and can be streamed directly without downloading:
+
+```python
+from datasets import load_dataset
+
+# Load particles (truth-level)
+particles_ds = load_dataset(
+    "OpenDataDetector/ColliderML_ttbar_pu0",
+    "particles",
+    split="train",
+    streaming=True
+)
+
+# Load tracker hits
+tracker_hits_ds = load_dataset(
+    "OpenDataDetector/ColliderML_ttbar_pu0",
+    "tracker_hits",
+    split="train",
+    streaming=True
+)
+
+# Load calorimeter hits
+calo_hits_ds = load_dataset(
+    "OpenDataDetector/ColliderML_ttbar_pu0",
+    "calo_hits",
+    split="train",
+    streaming=True
+)
+
+# Load reconstructed tracks
+tracks_ds = load_dataset(
+    "OpenDataDetector/ColliderML_ttbar_pu0",
+    "tracks",
+    split="train",
+    streaming=True
+)
+```
+
+### Example: Iterating Over Events
+
+```python
+import numpy as np
+
+# Iterate over first 10 events
+for i, event in enumerate(particles_ds.take(10)):
+    event_id = event['event_id']
+    n_particles = len(event['particle_id'])
+
+    print(f"Event {event_id}: {n_particles} particles")
+
+    # Access list columns as numpy arrays
+    px = np.array(event['px'])
+    py = np.array(event['py'])
+    pz = np.array(event['pz'])
+
+    # Compute transverse momentum
+    pt = np.sqrt(px**2 + py**2)
+    print(f"  Mean pt: {pt.mean():.2f} GeV")
+```
+
+### Example: Computing Track Features
+
+```python
+import numpy as np
+
+for event in tracks_ds.take(5):
+    # Get track parameters
+    qop = np.array(event['qop'])
+    theta = np.array(event['theta'])
+    phi = np.array(event['phi'])
+
+    # Compute derived quantities
+    pt = np.abs(1.0 / qop) * np.sin(theta)
+    eta = -np.log(np.tan(theta / 2.0))
+
+    print(f"Event {event['event_id']}: {len(qop)} tracks")
+    print(f"  pt range: [{pt.min():.2f}, {pt.max():.2f}] GeV")
+    print(f"  eta range: [{eta.min():.2f}, {eta.max():.2f}]")
+```
+
+### Example: Matching Tracks to Particles
+
+```python
+# Load both datasets
+particles = load_dataset("OpenDataDetector/ColliderML_ttbar_pu0", "particles", split="train", streaming=True)
+tracks = load_dataset("OpenDataDetector/ColliderML_ttbar_pu0", "tracks", split="train", streaming=True)
+
+# Process event-by-event
+for particle_event, track_event in zip(particles, tracks):
+    assert particle_event['event_id'] == track_event['event_id']
+
+    # Create particle ID lookup
+    particle_ids = np.array(particle_event['particle_id'])
+    particle_pt = np.sqrt(
+        np.array(particle_event['px'])**2 +
+        np.array(particle_event['py'])**2
+    )
+
+    # Get track associations
+    track_particle_ids = np.array(track_event['majority_particle_id'])
+
+    # Find matched particles
+    for track_idx, pid in enumerate(track_particle_ids):
+        if pid in particle_ids:
+            particle_idx = np.where(particle_ids == pid)[0][0]
+            truth_pt = particle_pt[particle_idx]
+
+            # Compute reconstructed pt
+            qop = track_event['qop'][track_idx]
+            theta = track_event['theta'][track_idx]
+            reco_pt = abs(1.0 / qop) * np.sin(theta)
+
+            print(f"Track {track_idx}: truth pt = {truth_pt:.2f}, reco pt = {reco_pt:.2f} GeV")
+```
+
+### Data Location
+
+The Parquet files are hosted at:
+
+```
+https://portal.nersc.gov/cfs/m4958/ColliderML/hard_scatter/ttbar/v1/parquet/
+├── truth/
+│   └── particles/
+│       ├── hard_scatter.ttbar.v1.truth.particles.events0-9.parquet
+│       ├── hard_scatter.ttbar.v1.truth.particles.events2000-2999.parquet
+│       └── ... (29 files total, ~29,000 events)
+├── reco/
+│   ├── tracker_hits/
+│   │   ├── hard_scatter.ttbar.v1.reco.tracker_hits.events0-9.parquet
+│   │   └── ... (29 files)
+│   ├── calo_hits/
+│   │   ├── hard_scatter.ttbar.v1.reco.calo_hits.events0-9.parquet
+│   │   └── ... (29 files)
+│   └── tracks/
+│       ├── hard_scatter.ttbar.v1.reco.tracks.events0-9.parquet
+│       └── ... (29 files)
+```
+
+### File Naming Convention
+
+Files follow the pattern:
+```
+<campaign>.<dataset>.<version>.<category>.<object>.<event_range>.parquet
+```
+
+Example: `hard_scatter.ttbar.v1.reco.tracks.events0-9.parquet`
+- Campaign: `hard_scatter`
+- Dataset: `ttbar`
+- Version: `v1`
+- Category: `reco` (or `truth`)
+- Object: `tracks`
+- Event range: `events0-9` (inclusive)
+
+### Performance Tips
+
+1. **Streaming**: Use `streaming=True` to avoid downloading the entire dataset
+2. **Batch processing**: Process events in chunks for better memory efficiency
+3. **Parallel loading**: Use `num_proc` parameter for multi-threaded data loading
+4. **Selective loading**: Only load the data types you need (particles, hits, tracks)
+
+### Related Datasets
+
+- **ColliderML_ttbar_pu200** (coming soon): Same process with 200 pileup interactions
+- **ColliderML_higgs_pu0** (coming soon): Higgs boson production without pileup
+
+### Support
+
+For questions, issues, or feature requests:
+- Open an issue on GitHub: https://github.com/ATLAS-ITk-ML/colliderml/issues
+- Email: danieltm@lbl.gov
+
+### Acknowledgments
+
+This work was supported by:
+- ATLAS ITk ML Reconstruction project
+- NERSC computing resources
+- U.S. Department of Energy, Office of Science
+
+---
+
+**Last updated**: October 2024
+**Dataset version**: v1