Memecoin Launch Trace Dataset

This repository contains the dataset and code for the submission: "MELT: A Behavioral Trace Dataset for High-Risk Memecoin Launch Detection"

Environment

• Python 3.9
• Install packages using:

pip install -r requirements.txt

Part 1: Feature Generation

Step 1: Download Dataset

Since the raw transaction data is very huge (>1TB), we only provide the parsed transaction datasets on Google Drive:

• pre_migration_tx.zip — Pre-migration (bonding curve) transactions. Required for feature generation. Download and unzip under MELT/data/tx/ (so the parsed transactions sit under data/tx/...).
• bundle.zip — Bundle trace data. Required for feature generation. Download and unzip under MELT/data/ (expands into data/bundle/).
• post_migration_tx.zip — Post-migration (Raydium DEX) transactions. Optional, very large, not used by feature generation. Only download if you want to do your own post-migration analysis.

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.

Step 2: Feature Generation

cd MELT/src
python feat_gen.py

This generates data/feat/feature.pkl from the pre-migration transactions, bundle trace data, and contextual information.

Part 2: High-risk Launch Detection

Step 1: Train a model

cd MELT/src
python train.py --model rf

--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:

flag	default	applies to
--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.

Step 2: Evaluate predictions

evaluate.py reads a prediction CSV from Step 1 and prints AUPRC plus a classification_report at one or more probability thresholds.

# single threshold
python evaluate.py --csv lgbm_pred_0.559999.csv --thresholds 0.5

# multi-threshold sweep
python evaluate.py --csv lgbm_pred_0.559999.csv --thresholds 0.3 0.4 0.5 0.6

flag	default	role
--csv	—	prediction CSV (relative paths resolve against results/)
--thresholds	[0.49]	one or more probability cutoffs; one classification_report per threshold

src/

file	role
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.

License

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.