View all files | ||||
This repository contains the dataset and code for the submission: "MELT: A Behavioral Trace Dataset for High-Risk Memecoin Launch Detection"
Since the raw transaction data is very huge (>1TB), we only provide the parsed transaction datasets on Google Drive:
You can skip Step 2 and download our pre-generated feature.pkl directly. Place it at MELT/data/feat/feature.pkl and proceed to Part 2. In this case neither pre_migration_tx.zip nor bundle.zip is needed.
This generates data/feat/feature.pkl from the pre-migration transactions, bundle trace data, and contextual information.
--model accepts any of: rf, xgb, lgbm, lr, mlp, or tcn, lstm, gru, transformer (time-series models). Prediction CSVs are written to MELT/results/{model}_pred_*.csv.
Common flags:
| --model | xgb | all |
| --epochs | 20 | DNN models |
| --batch_size | 256 | DNN models |
| --lr | 1e-3 | DNN models |
| --seed | 42 | all (Python random, numpy, torch, sklearn random_state, DataLoader shuffle) |
train.py only reports AUPRC (threshold-free) and dumps per-run prediction probabilities to results/. Threshold-based metrics (precision / recall / F1) and ensembling are done in the next step.
evaluate.py reads a prediction CSV from Step 1 and prints AUPRC plus a classification_report at one or more probability thresholds.
| --csv | — | prediction CSV (relative paths resolve against results/) |
| --thresholds | [0.49] | one or more probability cutoffs; one classification_report per threshold |
| feat_gen.py | Generates data/feat/feature.pkl from parsed transactions, bundle traces, and contextual info. |
| dataset.py | Data loading & preprocessing. Reads feature.pkl + label CSV, merges by mint_address, splits & scales. Exposes load_dataset(), TSDataset, ts_collate. |
| model.py | All model definitions and factories (sklearn baselines + MLP / TS deep models). |
| train.py | Training entry point. Argparse-driven; reports AUPRC and writes per-run prediction CSVs to results/. |
| evaluate.py | Evaluates prediction CSVs from train.py at one or more thresholds; supports weighted ensembling of multiple CSVs. |
This project is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. See LICENSE for the full text.
You are free to use, share, and adapt the material for non-commercial purposes, provided you give appropriate credit (please cite our paper). Commercial use requires separate permission from the authors.