K0d3r0x/dlgenai-nppe-dataset

NPPE2 T3 Training Logs Dataset (25-t3-nppe2.parquet)

Dataset Summary

This dataset contains logged training and validation metrics produced during experiments for the NPPE2 course
(Deep Learning & Generative AI, Term 3, 2025).

Each row corresponds to a single logging step from a model training run, capturing losses, accuracies, and F1 scores from models trained for protein secondary structure prediction using Q8 (8-state) and Q3 (3-state) labels.

This dataset does not contain protein sequences or biological samples.
It is intended for reproducibility, benchmarking, training-dynamics analysis, and educational use within the IITM Online Degree program.

---

Files in This Repository

• 25-t3-nppe2.parquet — NPPE2 training and validation metrics (primary dataset)
• 25-t3-nppe1.parquet — Earlier course phase metrics (archived for completeness)

Both files are stored at the repository root and can be loaded directly using Hugging Face Datasets.

---

Dataset Structure

• Format: Parquet
• Rows: 1,102
• Split: train

Note: The train split contains experiment log records, not samples used to train a model.

---

Features

Column	Type	Description
id	float64	Unique identifier for the logged record
timestamp	string	Time at which the metric was logged
run_name	string	Name of the experiment run
step	int64	Training step within the run
epoch	int64	Epoch number
train_loss	float64	Masked token-level training loss
q8_token_acc	float64	Token-level accuracy for Q8 during training
q3_token_acc	float64	Token-level accuracy for Q3 during training
q8_f1_micro	float64	Training-phase micro-averaged Q8 F1 score
q3_f1_micro	float64	Training-phase micro-averaged Q3 F1 score
harmonic_f1	float64	Harmonic mean of training Q8 and Q3 F1 scores
val_q8_f1	float64	Validation Q8 F1 score (legacy or non-micro logging)
val_q3_f1	float64	Validation Q3 F1 score (legacy or non-micro logging)
val_harmonic_f1	float64	Harmonic mean of validation Q8 and Q3 F1 scores
val_q8_f1_micro	float64	Validation micro-averaged Q8 F1 (Kaggle-aligned)
val_q3_f1_micro	float64	Validation micro-averaged Q3 F1 (Kaggle-aligned)
val_harmonic_f1_micro	float64	Validation harmonic micro F1 (primary selection metric)
f1_q8	float64	Final Q8 F1 score for the selected checkpoint
f1_q3	float64	Final Q3 F1 score for the selected checkpoint
harmonic_score	float64	Final harmonic score used for model comparison
hier_loss	float64	Hierarchical consistency regularization loss
q8_f1	float64	Q8 evaluation F1 (post-training evaluation)
q3_f1	float64	Q3 evaluation F1 (post-training evaluation)
harmonic	float64	Harmonic evaluation metric (post-training)

---

Usage

from datasets import load_dataset

ds = load_dataset(
    "K0d3r0x/dlgenai-nppe-dataset",
    data_files="25-t3-nppe2.parquet"
)

print(ds)
print(ds["train"][0])

# or you can also use this
# ds = load_dataset(
#     "K0d3r0x/dlgenai-nppe-dataset",
#     name="nppe2"
# )

# print(ds["train"].column_names)

---

Supported Use Cases

Analysis of training dynamics for sequence models

Benchmarking Q8/Q3 protein secondary structure models

Reproducibility of NPPE2 coursework experiments

Educational assignments and reporting

---

Notes

This dataset contains derived experiment logs only.

No biological sequences, personal data, or proprietary information are included.

Metrics align with Kaggle-style token-weighted evaluation where indicated.

---

Citation

If you use this dataset, please cite the IITM Online Degree NPPE2 course materials and acknowledge the dataset creator.