{"id":"a46dcf2b-dc0a-47be-8dca-761fd5603465","arxiv_id":"2411.10325","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A new public Bitcoin transaction graph dataset with 252M nodes, 785M edges, and ~34K labeled entities, plus GNN baselines.","lead":"The paper introduces a public Bitcoin transaction graph with 252 million nodes and 785 million edges spanning 13 years, along with roughly 34,000 entity labels. It also trains several graph neural networks as baselines, offering a large benchmark for cryptocurrency research.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Label accuracy from ChatGPT-based pipeline is unvalidated; without precision/recall estimates, the dataset's supervised benchmark value is unproven.","rationale":"The reader's weakest_assumption identifies the ChatGPT-based labeling pipeline as the most load-bearing premise. I concur. The paper's primary contribution is not just the raw transaction graph (which is plausibly constructed from the public blockchain) but the labeled subset that enables supervised tasks and benchmark comparisons. All downstream claims—GNN baselines, entity-type prediction, and the dataset's 'ground truth' utility—depend on the correctness of these labels. The paper provides no manual validation, no precision or recall, and its 'Limits' paragraph explicitly admits that forum content and ChatGPT inference are potential error sources. This is not merely a matter of 'outside current consensus'; it is an internal gap between the claimed benchmark quality and the evidence provided. Other possible concerns, such as the density figure (stated as ~1%, mathematically implausible for a graph with 785M edges and 252M nodes) and minor typographical errors, are secondary: they do not directly invalidate the central resource, though they should be corrected. The technical validation via GNN performance is circular as a validation of label quality, since there is no external gold standard and the models could fit systematic label noise. The proposed concrete test—manual annotation of a stratified random sample with inter-annotator agreement—would directly measure label reliability and settle whether the concern lands. If precision is high, the concern is resolved; if not, the dataset's supervised benchmark claims would need substantial qualification. Given that the dataset itself may still be valuable for unsupervised or structural analysis, a CONDITIONAL verdict (as the reader gave) is appropriate unless the validation fails outright, in which case REJECT or a strong qualification would be warranted. My reading does not change the reader's verdict.","tokens_in":12678,"tokens_out":3243,"duration_ms":32250,"concrete_test":"Independently construct a gold-standard test set: randomly sample 1,000 labeled addresses from the BitcoinTalk-derived portion (preserving category proportions). For each address, have two human annotators examine the original forum post plus the associated blockchain transaction context (and, for exchanges/marketplaces, cross-check official websites or known repositories such as WalletExplorer) to determine the true entity and entity type. Compute per-category precision, recall, and F1 for the ChatGPT-derived labels, as well as inter-annotator agreement (Cohen's kappa). If overall precision is below 90% or kappa is below 0.8, the labeling pipeline is too noisy to support the claim that the dataset serves as reliable ground truth for supervised learning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the dataset provides reliable ground-truth entity labels for 34,098 nodes and ~101k addresses—rests almost entirely on an automated labeling pipeline that uses ChatGPT (gpt-4o-mini) to extract addresses and infer entity types from BitcoinTalk posts, with no manual validation, precision, or recall figures. The paper's 'Limits' paragraph concedes that forum posts may contain misinformation and that ChatGPT has limitations in inferring entities from short unstructured context, but no quantitative assessment is given. Since 86.1% of labeled addresses come from BitcoinTalk, any systematic error in ChatGPT's extraction or classification (e.g., misidentifying a user's address as a service deposit address, or confusing entity categories) directly corrupts the supervised labels and undermines the GNN baselines and the dataset's use as a benchmark. The 'Technical Validation' section treats moderate macro-F1 scores (0.56–0.64) as evidence of label quality, but that reasoning is circular: models trained on noisy labels can still fit them, and no comparison to any external gold standard is provided. Without label-quality estimates, the dataset's headline utility for supervised graph learning remains unproven.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a large-scale Bitcoin transaction graph dataset, with 252 million nodes (clusters of locking scripts) and 785 million directed edges, spanning the first 700,000 blocks (~13 years) and 670 million transactions. Each node and edge carries temporal and aggregate features. Roughly 34,098 nodes are labeled by entity type, derived from about 101,186 labeled Bitcoin addresses. Address labels were obtained primarily by running ChatGPT (gpt-4o-mini) over BitcoinTalk forum posts to infer entity names and types, supplemented by several external sources (CoinMarketCap, DefiLlama, ransomware datasets, SDN list, etc.). The authors train GCN, GraphSage, GAT, GIN, and a gradient boosting classifier on a node classification task and report macro-F1 scores between 0.56 and 0.64 as baselines. The dataset and code are released publicly.","tokens_in":12836,"tokens_out":3104,"duration_ms":31078,"significance":"If the data and labels are reliable, this is a potentially valuable community resource: it appears to be the largest public Bitcoin transaction graph with temporal information, node/edge attributes, and entity-type labels, and it also bundles the raw BitcoinTalk posts used for labeling. The authors provide code for graph construction, labeling, and prediction, which supports reproducibility. The main significance hinges on whether the automatically generated labels can serve as ground truth for supervised graph learning and for downstream analyses such as entity interaction studies and temporal evolution analysis. The scale and the public availability alone make the dataset a candidate benchmark, but the lack of label-quality evidence currently leaves the central claim of a reliable supervised benchmark unproven.","major_comments":[{"comment":"The paper's central claim of providing a labeled dataset for supervised graph learning rests on the quality of the labels produced by the ChatGPT-based pipeline (Section 'Node labels', subsection 'ChatGPT' and 'Prompts'). The 'Limits' paragraph concedes that forum posts may contain misinformation and that ChatGPT has limitations in inferring entities from short unstructured context, yet no manual validation, precision/recall estimates, or agreement statistics are provided. The 'Technical Validation' section uses macro-F1 scores of GNNs (Table 5) as evidence that 'the dataset’s reliability and relevance' is established, but this reasoning is circular: models trained and evaluated on the same automatically generated labels can fit those labels without demonstrating their correctness. An external gold standard, a human-annotated sample, or at least a reported precision estimate on a stratified subset is needed to support the use of these labels as ground truth.","section":"Node labels / Limits / Technical Validation"},{"comment":"The statement in 'Data Records' that 'The graph’s density is approximately 1%' is incorrect by many orders of magnitude. With 252,219,007 nodes and 785,954,737 edges, the directed density is E/(V(V-1)) ≈ 1.2 × 10^-8, or about 0.0000012%, not 1%. This is not a minor typo: it materially misrepresents the graph's sparsity and may mislead users about the expected connectivity and the feasibility of algorithms. The sentence should be corrected, or the intended quantity (e.g., density of a sampled subgraph) should be defined explicitly.","section":"Data Records"},{"comment":"The chain from an address label to a node label involves two steps that are not quantitatively validated: (i) the address-to-script mapping and the clustering heuristics (ref. 8) that group scripts into nodes, and (ii) the resolution of conflicting labels within a cluster, where the paper states that 'no label was assigned' but gives no statistics on how often conflicts occur or how many otherwise-valid addresses are discarded. Since the labeled nodes are the core of the supervised benchmark, the paper should report the accuracy of the clustering heuristics on a known set of addresses (e.g., addresses of well-known services) and the number of clusters affected by label conflicts. Without these numbers, the reliability of the 34,098 labeled nodes is not established.","section":"Node labels / Graph Construction"}],"minor_comments":[{"comment":"The abstract states '33,000 nodes' and 'nearly 100,000 Bitcoin addresses', while the Data Records section reports 34,098 labeled nodes and 101,186 labeled addresses (Table 3). The counts should be harmonized, and the abstract should use the final figures.","section":"Abstract / Data Records"},{"comment":"In Table 1, the column labeled 'b' is described as 'Node alias of the sender', which appears to be a copy-paste error; it should read 'Node alias of the recipient' (or 'receiver').","section":"Table 1"},{"comment":"There are several typographical and grammatical errors that should be fixed, including 'econonomy' (Background), 'databse' (Usage Notes), and 'is the is the username' (Usage Notes).","section":"Background & Summary / Usage Notes"},{"comment":"The prompt templates are said to be available in the code repository, but the paper itself does not include any sample prompt or output. Including a short example in the appendix would help readers understand the extraction and inference process without needing to inspect the code.","section":"Node labels / Prompts"},{"comment":"The heuristics used for CoinJoin detection and colored-coin detection are cited from the first author's prior work (refs. 9 and 10, respectively). A brief description of the heuristics' expected false-positive/false-negative rates, or at least a statement of whether they have been validated independently, would increase confidence in the graph construction.","section":"Methods / Graph Construction"}],"recommendation":"major_revision","confidential_remarks":"The dataset is a substantial engineering contribution and the public release is commendable. My recommendation of major revision is driven by the unvalidated labeling pipeline, which is the load-bearing component for the supervised benchmark claim. The density error, while easily fixable, is a red flag for the rigor of the numerical reporting and should be corrected. I would not reject the paper, because the label-quality issue can in principle be addressed with a manual validation study or error-rate estimates, and the underlying graph data may still be valuable even with noisy labels if disclosed as such. I also note that the paper does not compare against the Elliptic datasets in the experiments, which would be a useful external reference point for the classification task."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline is the dataset: 252M nodes, 785M timestamped edges, built from the first 700,000 blocks with standard clustering heuristics, plus code and data released. That is a real contribution. Elliptic and Elliptic2 are smaller and binary-labeled; this is the first public graph at this scale with 13 years of temporal information and multi-class entity labels. The graph construction is careful—CoinJoin and colored-coin exclusions are sensible, and the node/edge schema is well documented. The GNN baselines are not novel, but they are competent and provide a reasonable starting point for future work. For a data descriptor, that is enough.\n\nThe soft spots are real but not fatal to the resource’s core value. The density claim is wrong by orders of magnitude: 785M edges in a graph with 252M nodes gives density around 1e-8, not 1%. That error needs correcting. The address counts also do not add up—the category table sums to ~103,600, while the text claims 101,186. These are the kind of sloppy numbers that erode trust in a dataset paper.\n\nThe bigger issue is the label set. 86% of addresses come from BitcoinTalk posts processed by gpt-4o-mini with no manual validation, no precision/recall, no error estimates. The paper’s own limits section concedes forum misinformation and ChatGPT’s limitations. The technical validation, which uses the labels as ground truth to train GNNs and reports macro-F1 around 0.6, is circular: it shows the models can fit the labels, not that the labels are correct. Without a randomly sampled human audit of a few hundred addresses, the supervised benchmark value is unproven. That said, the labels are clearly useful for exploratory analysis and for entity resolution heuristics; the risk is mainly for researchers who treat the labels as gold for training classifiers.\n\nWho is this for? Anyone working on Bitcoin graph analysis, anti-money-laundering, or large-scale temporal GNN benchmarks. The paper is worth sending to review, but a serious referee should require the factual corrections, a label-quality audit, and error bars on the baselines before acceptance. I would not desk-reject it; the dataset itself is significant and the flaws are fixable.","headline":"A genuinely large, usable Bitcoin transaction graph with a label set that is currently unvalidated; the resource deserves review, but the labeling pipeline needs real error analysis before it becomes a benchmark.","tokens_in":13383,"tokens_out":1841,"would_cite":false,"duration_ms":19072,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The authors release a 252-million-node, 785-million-edge temporal Bitcoin transaction graph claimed to be the largest public dataset of its kind, with entity labels and graph-neural-network baselines.","keywords":["Bitcoin","transaction graph","blockchain dataset","graph neural networks","entity labeling","temporal graph","language model annotation","cryptocurrency analysis"],"falsifier":"Take a random sample of, say, 200 labeled addresses from the BitcoinTalk-derived portion, have a human expert verify the entity name and type against the linked forum posts and on-chain behavior, and compute precision; if precision drops below a plausible threshold (e.g., 80%), or if the errors are systematically concentrated in a few entity types, the dataset's validity as ground truth for supervised learning is undermined.","tokens_in":12439,"feed_emoji":"📊","tokens_out":5021,"duration_ms":43054,"temperature":0.7,"pith_summary":"The paper introduces a large-scale temporal transaction graph of Bitcoin, built from the first 700,000 blocks, in which nodes are clusters of locking scripts (entities) and edges are value transfers between them. The authors claim this is the largest publicly available Bitcoin transaction dataset, with 252 million nodes and 785 million edges, spanning nearly 13 years and 670 million transactions. To make the graph useful for supervised learning, they label about 34,000 nodes by entity type using addresses harvested from BitcoinTalk forum posts via a large language model, plus supplementary official and research sources. They then train four graph neural networks and a gradient boosting classifier to predict node labels, establishing baseline macro-F1 scores up to 0.64. If the dataset is sound, it lowers the barrier to Bitcoin graph research by removing the need to build and cluster the blockchain graph from scratch.","feed_headline":"Largest public Bitcoin transaction graph arrives with 785M edges","feed_subtitle":"Temporal graph of 252M entities plus GNN baselines aims to lower the barrier to blockchain research.","key_machinery":"The central object is the transaction graph itself, built by a three-stage pipeline: (1) parse the Bitcoin blockchain into script-level transactions, (2) cluster locking scripts into entity nodes using heuristics from prior work, excluding CoinJoin and colored-coin transactions to avoid poisoning clusters and edges, and (3) define directed edges from net value flow between sender and recipient aliases. The labeling machinery is a set of ChatGPT prompts that process BitcoinTalk posts together with transaction and price context, returning reasoning, entity names, and addresses, which are then mapped to a taxonomy of eleven entity types (individual, mining, exchange, marketplace, gambling, bet, faucet, mixer, ponzi, ransomware, bridge). The same addresses are propagated to scripts and clusters, and clusters with conflicting labels are left unlabeled.","core_discovery":"The central claim is that this is the largest publicly available Bitcoin transaction graph dataset designed for research, overcoming limitations of existing datasets such as Elliptic 1 and 2. The graph represents real entities rather than raw addresses: nodes are clusters of locking scripts grouped by heuristics that identify likely co-ownership, and directed edges represent value transfers computed from net value flows. Each node and edge carries a block-index timestamp, making the graph temporal. A labeling pipeline uses the language model ChatGPT to extract Bitcoin addresses from BitcoinTalk posts and infer the entity name and type, supplemented by address lists from exchanges, ransomware researchers, the U.S. sanctions list, mining coinbase messages, and the Wrapped Bitcoin bridge; nodes are labeled when their constituent scripts match labeled addresses and conflicts are resolved by discarding ambiguous clusters. As technical validation, node classification experiments show that graph neural networks (GAT and GIN) reach macro-F1 scores around 0.63-0.64, outperforming a gradient boosting classifier on tabular features, indicating that node features and graph structure carry predictive signal.","pith_inferences":["The labeling approach, if validated, implies that large language models can be used to semi-automatically build entity-identity ground truth from community forums, a method transferable to other blockchain ecosystems or other domains with public discussion and pseudonymous identifiers.","The paper's claim of being 'largest' depends on its definition of publicly available and graph-form; a natural test is a systematic comparison with other released datasets' sizes and label quality, which the paper does not perform.","Because the labels are derived from the same BitcoinTalk posts, there may be a temporal selection bias: entities active after the rise of LLM-era posts or those that announce themselves on forums are over-represented, affecting class distributions.","The confusion of many classes toward 'individual' suggests that the frontier between individual users and organizational addresses is fundamentally fuzzy, which could motivate multi-label or hierarchical labeling in future versions."],"forward_implications":["Researchers can directly benchmark graph learning methods on a temporal Bitcoin graph without building their own blockchain parser and clustering pipeline.","The temporal edge/node timestamps enable studies of network evolution, entity interaction dynamics, and event-driven behavior over nearly 13 years.","The 101k labeled addresses and 34k labeled nodes provide supervised signals for entity-type classification, and the 14M BitcoinTalk posts are included as a text corpus.","The dataset's scale (252M nodes, 785M edges) positions it as a pre-training resource for transaction-graph representation learning that could transfer to other payment networks.","The baseline results suggest that graph structure helps entity-type prediction, with GNNs outperforming tabular classifiers on several classes."],"supporting_citations":[{"why":"Defines the Bitcoin protocol and transaction model that the dataset is built on.","marker":"[1]"},{"why":"Provides the Elliptic 1 dataset that this work extends and compares against in scale.","marker":"[6]"},{"why":"Provides the Elliptic 2 dataset, the previous largest graph-based Bitcoin dataset.","marker":"[7]"},{"why":"Supplies the heuristic clustering method that maps scripts to entity nodes.","marker":"[8]"},{"why":"Supplies CoinJoin detection heuristics used to filter transactions that would corrupt clusters.","marker":"[9]"},{"why":"Establishes the language-model capability that the labeling pipeline relies on.","marker":"[11]"},{"why":"Provides ransomware addresses that are integrated as a label source.","marker":"[4]"},{"why":"Provides additional ransomware addresses used to enrich the label set.","marker":"[5]"}],"fun_headline_variants":["Largest public Bitcoin transaction graph: 785M edges","Bitcoin graph dataset: 252M nodes, 785M edges, 13 years","13-year Bitcoin transaction graph with GNN baselines","Public Bitcoin graph: 252M nodes, labeled, GNN baselines","Largest temporal Bitcoin graph: 252M nodes, 785M edges"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The dataset's supervised value rests on the assumption that labels extracted by the language model and mapped through the pipeline are accurate, yet no manual validation, precision, or recall estimate is provided for the ChatGPT-based labeling.","fun_headline_variants_meta":{"raw":{"variants":["Largest public Bitcoin transaction graph: 785M edges","Bitcoin graph dataset: 252M nodes, 785M edges, 13 years","13-year Bitcoin transaction graph with GNN baselines","Public Bitcoin graph: 252M nodes, labeled, GNN baselines","Largest temporal Bitcoin graph: 252M nodes, 785M edges"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000546,"raw_usage":{"total_tokens":2616,"prompt_tokens":955,"completion_tokens":1661,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":1566}},"tokens_in":571,"tokens_out":1661,"duration_ms":12103,"temperature":1.0,"reasoning_tokens":1566,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:44:25.961668+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of, say, 200 labeled addresses from the BitcoinTalk-derived portion, have a human expert verify the entity name and type against the linked forum posts and on-chain behavior, and compute precision; if precision drops below a plausible threshold (e.g., 80%), or if the errors are systematically concentrated in a few entity types, the dataset's validity as ground truth for supervised learning is undermined.","supporting_citations":[{"cited_title":"Bitcoin: A peer-to-peer electronic cash system","cited_arxiv_id":null,"evidence_quote":"Defines the Bitcoin protocol and transaction model that the dataset is built on."},{"cited_title":"& Ruj, S","cited_arxiv_id":null,"evidence_quote":"Provides ransomware addresses that are integrated as a label source."},{"cited_title":"& Dupont, B","cited_arxiv_id":null,"evidence_quote":"Provides additional ransomware addresses used to enrich the label set."}],"review_version":1}