TM-RugPull: Temporal Rug Pull Benchmark
- TM-RugPull is a benchmark dataset designed for causally valid, early detection of rug pulls using multimodal features and cross-chain data.
- It employs strict temporal hygiene by extracting features only from the first half of each project’s lifespan to prevent post-collapse leakage.
- Covering over 1,000 token projects across various categories, TM-RugPull addresses limitations in earlier DeFi-only datasets and enhances fraud detection.
Searching arXiv for TM-RugPull and closely related rug-pull detection/dataset papers to ground the article in current literature. A useful way to understand TM-RugPull is as a benchmark design for causally valid early-warning research on rug pulls in tokenized ecosystems. Introduced as “TM-RugPull: A Temporally Sound, Multimodal Dataset for Early Detection of Rug Pulls Across the Tokenized Ecosystem”, it is a public dataset of 1,028 token projects spanning Ethereum, BSC, Polygon, Arbitrum, and Fantom from 2016 to 2025, with features extracted strictly from the first half of each project’s lifespan and labels grounded in forensic corroboration and expert review (Shoaei et al., 25 Feb 2026). Its central purpose is to address persistent weaknesses in prior resources—especially temporal leakage, DeFi-only scope, on-chain-only feature design, and ambiguous labels—by providing a leakage-resistant, multimodal benchmark for studying rug-pull dynamics before collapse becomes obvious (Shoaei et al., 25 Feb 2026).
1. Definition, scope, and benchmark rationale
TM-RugPull is explicitly framed as a response to a methodological problem in rug-pull research: many existing datasets are not suitable for realistic early detection because they incorporate post-collapse evidence, constrain analysis to narrow on-chain views, focus mainly on DeFi, or rely on weak heuristics and non-public labeling procedures (Shoaei et al., 25 Feb 2026). The dataset is positioned not merely as another collection of scam tokens, but as a benchmark built to reflect what an analyst could have known before a scam fully unfolded (Shoaei et al., 25 Feb 2026).
The dataset’s coverage is broader than earlier DeFi-centric resources. It includes DeFi, meme coins, NFT/gaming projects, and celebrity-themed or impersonation-based tokens, rather than only protocol-style DeFi scams (Shoaei et al., 25 Feb 2026). This breadth is empirically motivated: the paper reports that non-DeFi tokens account for over 40% of scam cases in the dataset, indicating that DeFi-only benchmarks omit a substantial part of the attack surface (Shoaei et al., 25 Feb 2026). This suggests that TM-RugPull should be read as a cross-ecosystem benchmark rather than a protocol-specific fraud corpus.
The benchmark is public and released as RugPull1K on Hugging Face Datasets, under CC BY 4.0, with rugpull1k.csv, a data dictionary in features.md, and Python scripts for replication and visualization (Shoaei et al., 25 Feb 2026). The paper treats this release as the public instantiation of TM-RugPull, although it does not explicitly explain the naming distinction between TM-RugPull and RugPull1K (Shoaei et al., 25 Feb 2026).
2. Operational definition of a rug pull and labeling policy
TM-RugPull adopts a conservative operational definition rather than treating all suspicious short-lived tokens as scams. In the paper’s background formulation, a rug pull requires all three of the following: (i) an abrupt and near-total withdrawal of liquidity from decentralized exchanges, (ii) cessation of meaningful on-chain activity within a short post-withdrawal window, and (iii) collapse of token price or trading volume to a non-determinable or negligible state, leaving the asset effectively worthless (Shoaei et al., 25 Feb 2026).
In dataset construction, this is tightened into a concrete labeling rule. A project is labeled as a rug pull if and only if all three conditions hold for at least 72 consecutive hours. Those conditions are stated as zero liquidity, zero meaningful transactional activity, and undefined or untraceable price/volume on major DEXs (Shoaei et al., 25 Feb 2026). This rule makes the benchmark deliberately precision-oriented: suspicious behavior alone is insufficient, and ambiguous or borderline failures are excluded rather than folded into the positive class (Shoaei et al., 25 Feb 2026).
Labeling is also evidence-driven. A project is labeled scam/rug pull only when it both satisfies the formal operational rule and is corroborated by at least two independent sources, with CertiK, Rekt.news, and De.Fi named as examples (Shoaei et al., 25 Feb 2026). A project is labeled legitimate only if it shows more than one year of sustained on-chain activity and no scam reports over that period (Shoaei et al., 25 Feb 2026). Ambiguous cases are excluded entirely, reviewed by at least two analysts, and final labeling requires consensus (Shoaei et al., 25 Feb 2026).
This conservative design has methodological consequences. It reduces weak-label contamination and hindsight-driven categorization, but it also biases the benchmark toward clear-cut cases. A plausible implication is that TM-RugPull prioritizes analytical clarity over exhaustive ecological coverage of gray-area fraud.
3. Temporal hygiene and leakage resistance
The dataset’s core design principle is what the paper repeatedly calls temporal hygiene or temporal integrity (Shoaei et al., 25 Feb 2026). TM-RugPull introduces an explicit temporal anchor, the project midpoint, defined as the halfway point between project initiation and project termination or last observed activity (Shoaei et al., 25 Feb 2026). Every feature is extracted strictly before this midpoint, and all modalities are temporally aligned at the project midpoint (Shoaei et al., 25 Feb 2026).
This midpoint policy is meant to prevent contamination from post-collapse evidence such as the final price crash, the liquidity drain itself, or social-media backlash after exposure (Shoaei et al., 25 Feb 2026). The paper visually validates the policy with a midpoint-ratio figure centered near 0.5, which it uses to argue that the boundary is applied consistently across projects with heterogeneous lifespans (Shoaei et al., 25 Feb 2026).
The temporal design is one of TM-RugPull’s strongest contributions, but it is also bounded. The paper does not provide formal mathematical notation for midpoint computation, temporal window construction, or train/test chronology, and it does not describe a chronological train/test split for benchmark experiments (Shoaei et al., 25 Feb 2026). Thus TM-RugPull is leakage-resistant at the feature extraction level, but the paper stops short of defining a full calendar-based evaluation protocol.
This temporal framing closely aligns with later leakage-aware work. The LROO Rug Pull Detector paper similarly treats early rug-pull detection as a pre-collapse forecasting problem, using features extracted strictly prior to any liquidity withdrawal and evaluating on a temporally isolated 20\% test set (Shoaei et al., 11 Mar 2026). Taken together, these papers indicate a shift from retrospective scam classification toward causal forecasting under explicit temporal constraints (Shoaei et al., 11 Mar 2026).
4. Data collection, curation, and multimodal schema
TM-RugPull is built through a multi-stage workflow combining on-chain behavioral data, smart contract metadata, off-chain OSINT signals, and label evidence under a unified temporal framework (Shoaei et al., 25 Feb 2026). All projects are sourced from real deployed smart contracts on public blockchains, and the dataset contains no synthetic, simulated, or artificially generated data (Shoaei et al., 25 Feb 2026).
The curation process is unusually stringent. All 1,028 token projects were individually inspected by a team of three blockchain security analysts with expertise in DeFi, NFTs, and token forensics (Shoaei et al., 25 Feb 2026). The authors excluded projects with incomplete on-chain records, inaccessible contracts, inconsistent metadata, or ambiguous behavioral patterns, reporting that this filtering step removed over 200 borderline candidates (Shoaei et al., 25 Feb 2026). This reinforces the benchmark’s emphasis on precision and reproducibility.
The schema is organized into four conceptual categories (Shoaei et al., 25 Feb 2026):
| Category | Representative contents | Intended role |
|---|---|---|
| On-chain behavioral | token concentration metrics, holder distribution variance, transaction activity statistics, quarterly price indicators | structural manipulation and early market dynamics |
| Smart contract metadata | blockchain platform, consensus mechanism, token standard, contract deployment status | project and execution context |
| OSINT signals | Google Search hits/search visibility, Twitter/X activity/volume, social media activity indicators | hype, promotion, and off-chain precursors |
| Temporal anchors | project lifecycle timestamps, midpoint boundary, label provenance information | temporal alignment and provenance |
The paper does not provide a total feature count, a full field-by-field schema, or per-column descriptive statistics (Shoaei et al., 25 Feb 2026). However, it highlights representative variables such as token concentration ratio, holder variance among the top 1% holders, and quarterly price dynamics on the on-chain side, and Google Search hits and Twitter/X activity on the OSINT side (Shoaei et al., 25 Feb 2026). A central claim is that these modalities are not simply concatenated but temporally aligned and capped at the midpoint.
This multimodal design differentiates TM-RugPull from earlier code-only and transaction-only approaches. CRPWarner is a static-analysis system for contract-related rug pulls with strong code semantics but no OSINT layer (Lin et al., 2024). RPHunter fuses contract risk and transaction behavior through graph learning, but it remains centered on code-and-transaction fusion rather than temporally aligned OSINT (Wu et al., 23 Jun 2025). TM-RugPull occupies a complementary position: it is a dataset whose novelty lies in the combination of cross-category token diversity, multimodal evidence including OSINT, and strict temporal hygiene (Shoaei et al., 25 Feb 2026).
5. Empirical observations and illustrative benchmarking
The descriptive analysis reported for TM-RugPull identifies several characteristic differences between scam and legitimate projects. First, as noted above, over 40% of scam cases are outside DeFi, supporting the decision to include meme, NFT, and celebrity-themed tokens (Shoaei et al., 25 Feb 2026). Second, scam tokens show significantly higher token concentration and higher holder variance among top 1% holders than legitimate projects (Shoaei et al., 25 Feb 2026). In the conclusion, the paper specifies that scam tokens exhibit significantly higher token concentration with under a Mann–Whitney U test (Shoaei et al., 25 Feb 2026).
Third, the temporal alignment analysis suggests that OSINT activity peaks well before the midpoint while price remains stable, which the authors interpret as evidence of an authentic early-warning window (Shoaei et al., 25 Feb 2026). In the sample illustration, Twitter volume and Google search interest surge during Q1–Q2 before overt on-chain collapse (Shoaei et al., 25 Feb 2026). This suggests that off-chain hype may precede visible on-chain deterioration, providing a concrete rationale for multimodal early detection.
Benchmarking in the paper is explicitly illustrative rather than exhaustive (Shoaei et al., 25 Feb 2026). The authors compare models trained on the full benchmark against models trained on a DeFi-only subset, while keeping feature definitions, temporal constraints, labeling protocol, and preprocessing fixed (Shoaei et al., 25 Feb 2026). They use lightweight tabular classifiers, specifically logistic regression and random forest, not to establish model superiority but to isolate the contribution of cross-domain diversity (Shoaei et al., 25 Feb 2026).
The main result is qualitative: models trained on the full benchmark show clearer separation between scam and legitimate tokens, with less overlap near the decision boundary, whereas models trained on the DeFi-only subset show more overlap and greater uncertainty (Shoaei et al., 25 Feb 2026). The paper interprets this as evidence that heterogeneous token categories and temporally aligned multimodal signals improve separability and reduce overfitting to DeFi-specific artifacts (Shoaei et al., 25 Feb 2026). It further claims a significant decrease in false positives, though it does not provide standard benchmark metrics such as AUROC, F1, precision, recall, PR-AUC, or confusion matrices (Shoaei et al., 25 Feb 2026).
A later modeling paper operationalizes a closely related design philosophy at the detector level. The LROO Rug Pull Detector integrates on-chain and OSINT signals under strict temporal constraints and reports strong discriminative performance with TabPFN, including 0.982 accuracy, 0.982 F1, 0.997 ROC AUC, and 0.997 PR AUC on a temporally isolated 20\% test set (Shoaei et al., 11 Mar 2026). Because that framework is described as built around the same design direction later referenced as TM-RugPull, it can be read as a modeling blueprint for how a leakage-resistant multimodal benchmark can be operationalized (Shoaei et al., 11 Mar 2026).
6. Position within the rug-pull literature
TM-RugPull sits at the intersection of several strands of rug-pull research. Earlier work such as “Do not rug on me: Zero-dimensional Scam Detection” built pre-rug classifiers for Uniswap V2 using token propagation, liquidity-pool, and smart-contract heuristics, but within a narrower Ethereum/Uniswap setting (Mazorra et al., 2022). “Token Spammers, Rug Pulls, and SniperBots” provided a large-scale analysis of 1-day rug pulls on Ethereum and BSC, emphasizing disposable-token campaigns, creator concentration, and LP-event structure (Cernera et al., 2022). NFT-focused work has instead emphasized creator networks, contract backdoors, or marketplace behavior, as in “Understanding Rug Pulls: An In-Depth Behavioral Analysis of Fraudulent NFT Creators” (Sharma et al., 2023) and “A Deep Dive into NFT Rug Pulls” (Huang et al., 2023).
On the code-analysis side, CRPWarner formalized contract-related rug pulls through malicious functions such as Hidden Mint Function, Limiting Sell Order, and Leaking Token, evaluated via bytecode-level static analysis (Lin et al., 2024). NFT contract-focused static analysis later catalogued hidden backdoors such as selfdestruct, delegatecall, unrestricted minting, and owner-only withdrawal patterns in verified NFT contracts (Pathade et al., 9 Jun 2025). More recent multimodal work, especially RPHunter, has fused semantic code graphs and transaction-behavior graphs to capture the interplay between malicious token logic and suspicious early transaction evolution (Wu et al., 23 Jun 2025).
TM-RugPull’s distinctiveness lies in benchmark design rather than detector architecture. The paper presents it as the first public dataset to combine multi-chain coverage, cross-category token diversity, multimodal features including OSINT, strict temporal hygiene, and manually verified labels grounded in forensic evidence (Shoaei et al., 25 Feb 2026). In that sense, its main contribution is not only broader coverage but a more principled notion of what an early-warning benchmark should look like.
The dataset also highlights a tension in the literature between breadth and label sharpness. Systems like SolRPDS provide very large liquidity-centric datasets for Solana, but explicitly note that inactivity is a signal for suspicious behavior and not definitive evidence for rug pull (Alhaidari et al., 6 Apr 2025). TM-RugPull instead sacrifices scale to preserve stronger label semantics and human verification (Shoaei et al., 25 Feb 2026). This suggests a methodological divide between large proxy-label corpora and smaller, forensic-grade benchmarks.
7. Limitations, naming issues, and future directions
The paper does not collect its limitations into a dedicated section, but several are clear. Coverage is restricted to five blockchains—Ethereum, BSC, Polygon, Arbitrum, and Fantom—so TM-RugPull is not comprehensive across all token ecosystems (Shoaei et al., 25 Feb 2026). Because labels require strong forensic corroboration and long observation periods for legitimate tokens, the benchmark may underrepresent borderline or emerging scams (Shoaei et al., 25 Feb 2026). The exclusion of over 200 ambiguous candidates improves precision but narrows ecological coverage (Shoaei et al., 25 Feb 2026).
The use of OSINT signals such as Google Search and Twitter/X introduces possible issues of platform noise, manipulation, and measurement instability, though the paper does not elaborate them in detail (Shoaei et al., 25 Feb 2026). The legitimacy rule requiring more than one year of sustained on-chain activity may also induce a form of survivorship bias by making the negative class cleaner than what real-time deployment would face (Shoaei et al., 25 Feb 2026). Finally, despite the emphasis on temporal soundness, the paper does not define a formal chronological benchmark protocol, leaving future work to specify calendar-based train/test generalization (Shoaei et al., 25 Feb 2026).
A mild naming inconsistency also remains unresolved. The title and body call the benchmark TM-RugPull, while the data-availability section names the release RugPull1K (Shoaei et al., 25 Feb 2026). The paper appears to treat them as the same benchmark, but it does not explicitly explain the distinction.
Future work is strongly implied rather than fully specified. The most immediate extensions would be more formal temporal evaluation protocols, richer schema statistics, modality ablations, and broader chain coverage. A plausible implication is that TM-RugPull will be most influential not merely as a dataset to train on, but as a template for how rug-pull benchmarks should be constructed: temporally bounded, multimodal, forensically labeled, cross-ecosystem, and public.