{"id":"973ea51d-595c-419d-8068-0ffab97457ad","arxiv_id":"2509.05792","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"TPCpp-10M packages 10 million simulated sPHENIX TPC collision events and 90,000 labeled events in NumPy arrays for downstream ML benchmark tasks.","lead":"This paper releases a dataset of 10 million simulated proton-proton collisions recorded as time projection chamber spacepoints, plus labeled events for tracking, particle-ID, and noise-tagging benchmarks. The dataset is offered as a common resource for training machine learning foundation models in nuclear physics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'noise' labels are defined as pT<60 MeV/c truth-associated tracks, not detector noise; the noise-tagging benchmark is therefore mislabeled and one of the three claimed downstream tasks is not well defined.","rationale":"The reader's weakest assumption was the lack of validation against real sPHENIX data. That is a real limitation but not the most load-bearing issue for the dataset's central claim: the data are explicitly simulated, and the authors disclose that residual differences may exist. A stronger and more specific problem is internal: the noise-tagging label definition contradicts the task definition. Since the dataset's value partly rests on '70,000 labeled examples spanning three well-defined downstream tasks,' a mislabeled benchmark task directly undermines one of the three advertised tasks. This is a correctness flaw, not a disagreement with external consensus. I considered the label-count inconsistency (70k vs 90k) and unreported event-filtering rejections; these are real documentation issues but secondary. The noise-tag issue can be resolved by the proposed data inspection, and it does not necessarily invalidate the unlabeled 10M-event resource. Therefore the reader's CONDITIONAL verdict is appropriate and unchanged.","tokens_in":6138,"tokens_out":9630,"duration_ms":112678,"concrete_test":"Download the labeled test shard and cross-tabulate noise_tags with track_ids. If every spacepoint with noise_tags==1 has a valid (non-null) track_id, then the 'noise' label is assigned to truth-associated signal, not detector noise. Additionally, recover the pT of the associated truth tracks (from the generation-level truth or the provided per-track kinematics) and verify that the noise class is exactly pT<60 MeV/c; if so, the noise-tagging task is a kinematic threshold tag, not noise rejection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The labeled subset's noise-tagging task is internally inconsistent with its own label definition. In 'Data Description', noise_tags is defined as 1 iff the matched track's pT < 60 MeV/c and 0 otherwise. This is a kinematic cut on truth-matched tracks, not a detector-noise flag. Actual TPC noise hits (electronics pickup, unmatched clusters, etc.) have no matched track and are not represented by this rule. Moreover, a 60 MeV/c track produced at/near the event vertex has gyroradius ~14 cm in the 1.4 T field, below the TPC inner radius of ~32 cm, so such a primary would not reach the TPC active volume at all; the tagged hits must be secondaries or conversions, which are still real ionization. The downstream task description says noise tagging is 'binary classification noting whether or not a spacepoint is noise,' but the provided label cannot distinguish detector noise from signal. Thus one of the three benchmark tasks is not what it claims to be. This is a more immediate correctness flaw than the absence of real-data validation, because it can be settled directly from the released data and code. A corrected benchmark would require either genuine unmatched/noise hits or a renamed task (e.g., 'low-pT track tagging').","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TPCpp-10M, a dataset of 10 million simulated proton-proton collisions at sqrt(s)=200 GeV for the sPHENIX TPC, generated with Pythia8 and Geant4 through the sPHENIX software stack. The dataset is released in NumPy format, with an unlabeled set for self-supervised pretraining and a labeled subset for track finding, particle identification, and noise tagging. The authors provide repository URLs, commit hashes, and an sPHENIX build release number for reproducibility, and they recommend evaluation metrics for each downstream task. The central claim is that this is a large, openly accessible, high-fidelity TPC dataset that lowers the barrier for machine learning researchers to work on particle tracking.","tokens_in":6467,"tokens_out":2971,"duration_ms":34743,"significance":"If the dataset is indeed of the claimed size and quality, it fills a real gap: most public tracking datasets are silicon-based, whereas TPC data have different pattern-recognition challenges (long drift, diffusion, many layers). The 10M-event scale is well suited for self-supervised foundation-model training, and the NumPy format plus example scripts make it accessible outside the particle-physics community. The reproducibility measures are a strong point: commit hashes plus the CVMFS release (ana.435) give a concrete, auditable generation pipeline. The three labeled tasks are sensible in principle, and the double-majority tracking metric is a standard choice. The main value is the resource itself; however, several specific label and consistency issues in the labeled subset must be addressed before the benchmark claims are reliable.","major_comments":[{"comment":"The labeled-set size is internally inconsistent. The abstract and Value of the Data state 70,000 labeled events, but the Data Description (and Figure 3) list 7 train shards of 10k events each (70k), plus a 13k validation file and a 7k test file, totaling 90k. This discrepancy affects every downstream benchmark comparison, because readers cannot know the actual train/validation/test sizes or whether the abstract is undercounting. Please state exactly how many events are in each split and correct the abstract/value statement accordingly.","section":"Data Description / labeled folder structure"},{"comment":"The noise-tagging task is mislabeled. noise_tags is defined as 1 iff the matched track's pT < 60 MeV/c and 0 otherwise, and the task description says it flags noise. This is a kinematic cut on truth-matched tracks, not a detector-noise flag. Real TPC noise (electronics pickup, unmatched clusters, random combinations) has no truth track and is not captured by this rule. The stated rationale that such tracks are 'unlikely to reach the TPC active volume' is also quantitatively odd: a 60 MeV/c primary in the 1.4 T field has gyroradius ~14 cm, well below the TPC inner radius of ~32 cm, so it cannot reach the active volume at all; the tagged particles must be secondaries or conversions. One of the three downstream tasks is therefore not what it claims to be. Either rename the task to 'low-pT track tagging' or provide genuine unmatched/noise hits as the positive class.","section":"Data Description, noise_tags definition; Figure 5(b)"},{"comment":"There is no validation against real sPHENIX collision data. The manuscript claims 'realistic detector conditions' and 'high-fidelity resource,' but no comparison is shown of spacepoint multiplicities, hit resolutions, dE/dx distributions, or tracking performance between simulation and real TPC data. The Limitations section only says 'Residual differences from real data may exist.' For a dataset whose primary purpose is to stand in for real data in foundation-model training, this is a load-bearing gap. Please add at least basic quantitative cross-checks (e.g., spacepoint distributions, cluster shapes, or tracking efficiency against available sPHENIX p+p reference data) or explicitly qualify the dataset as simulation-only with no fidelity claim.","section":"Limitations / Experimental Design"},{"comment":"The preprocessing section states that tracks with fewer than five spacepoints are removed and events with fewer than 20 remaining spacepoints are dropped. It is not stated whether these cuts are applied to the unlabeled 10M set, the labeled 90k set, or both. If the labeled noise-tagging task is meant to include detector noise, removing short tracks and requiring every spacepoint to belong to a truth track would already eliminate the very objects that should be labeled as noise. Clarify the filter scope and report how many events were rejected at each step.","section":"Data preprocessing / labeled subset applicability"}],"minor_comments":[{"comment":"The mismatch between 70k and 90k should be fixed everywhere it appears; see major comment.","section":"Abstract / Data Description"},{"comment":"Even after renaming, the phrase 'unlikely to reach the TPC active volume' should be replaced with a precise statement about production radius and pT thresholds, since a primary track at 60 MeV/c cannot reach the TPC at all.","section":"Data Description, noise_tags wording"},{"comment":"Minor typo: 'detector and1.4 Tmagnetic field' is missing a space after 'and'.","section":"Experimental Design"},{"comment":"The caption repeats 'from fewer than 10 to nearly 100 truth tracks per event' twice; the second occurrence should be removed or reworded.","section":"Figure 4 caption"},{"comment":"The double-majority matching section would benefit from a one-sentence note on how to treat candidates/particles with zero hits in the denominator, though this is standard practice and not blocking.","section":"Evaluation Metrics"}],"recommendation":"major_revision","confidential_remarks":"The resource is potentially useful and the reproducibility measures are commendable. The main issue is not the absence of real-data validation per se, but the fact that the noise-tagging label is not detector noise and the labeled-set size is inconsistent. These are fixable within the manuscript's scope by renaming the task, adding true noise hits, and correcting the counts. I would not reject, but the current version cannot be accepted as a benchmark resource because the benchmark descriptions are misleading. The editor may also want to ensure the dataset DOI actually contains the files matching the claimed split sizes before final acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely useful dataset resource, and the main thing you should know is that one of its three benchmark tasks—noise tagging—is not what it claims to be. The 'noise' label is a kinematic cut (pT < 60 MeV/c) on truth-matched tracks, not detector noise. That's a real flaw, but it's fixable, and the rest of the resource holds up.\n\nWhat's actually new: it's the first large (10M event) TPC-specific simulated tracking dataset in a common NumPy format with truth associations, generated with the sPHENIX software stack (Pythia8 + Geant4) and pinned commit hashes plus a Zenodo DOI. That fills a real gap: most public tracking datasets are silicon-based, and TPC-specific features like long drift and diffusion are otherwise hard for ML researchers to get without running their own simulation. The labeled subset, despite the noise-label problem, does provide genuine track IDs and PID labels, and the recommended double-majority metrics for track finding are sensible. The documentation of event complexity distributions and class ratios is also helpful.\n\nSoft spots, in order. First, the noise tagging task. The paper defines noise_tags as 1 if the matched track's pT < 60 MeV/c. As you'd expect, a 60 MeV/c primary in the 1.4 T field has a gyroradius of ~14 cm, below the TPC inner radius of ~32 cm, so it would never reach the active volume. The tagged hits are presumably secondaries or conversions—still real ionization, not detector noise. The task description says 'binary classification noting whether or not a spacepoint is noise,' which the label cannot support. This is not a philosophical quibble; either redefine the task or provide actual unmatched/noise hits. Second, the labeled-set count is inconsistent: the abstract says 70k, but train (70k) + validation (13k) + test (7k) sums to 90k. Likely a typo, but it matters for benchmarking. Third, the preprocessing removes tracks with fewer than five spacepoints and rejects events with fewer than 20 remaining spacepoints, but the fraction of events rejected isn't reported. For a benchmark resource, that number matters. The lack of validation against real sPHENIX data is a limitation, but for an ML-focused simulation resource it's not a blocker.\n\nBottom line: this deserves a serious referee. The central resource is valuable and the reproducibility infrastructure is strong. The noise-tag label and the count inconsistencies need to be fixed before the benchmark tasks are used as-is. I'd recommend conditional acceptance with those changes.","headline":"Useful public TPC simulation dataset, but its 'noise tagging' benchmark is mislabeled and needs a fix before serious use.","tokens_in":6945,"tokens_out":2192,"would_cite":true,"duration_ms":22830,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"TPCpp-10M provides 10 million simulated proton–proton collisions as TPC spacepoints in a common array format, with 70,000 labeled events for track finding, particle identification, and noise tagging.","keywords":["time projection chamber","foundation models","self-supervised learning","track finding","particle identification","noise tagging","benchmark dataset","simulated collisions"],"falsifier":"Compare the spacepoint multiplicity, noise occupancy, and hit-position residuals in this dataset against real minimum-bias proton–proton data from the same TPC at 200 GeV; a large disagreement in any of these distributions would break the claim that the simulation is a faithful surrogate.","tokens_in":6077,"feed_emoji":"⚛️","tokens_out":8985,"duration_ms":89387,"temperature":0.7,"pith_summary":"TPCpp-10M is a public dataset of 10 million simulated proton–proton collisions at 200 GeV, stored as time projection chamber (TPC) spacepoints in a common array format. Ten million unlabeled events are intended for self-supervised pretraining of machine-learning foundation models, while 70,000 labeled events carry ground-truth labels for track finding, particle identification, and noise tagging. The dataset is designed to give machine-learning researchers a low-barrier, reproducible testbed for TPC tracking, complementing public tracking datasets that come mostly from silicon detectors. The paper's central claim is that this resource is large and faithful enough to support scaling studies and transfer of foundation models to nuclear and particle physics.","feed_headline":"10M simulated proton collisions released as an AI training dataset","feed_subtitle":"Unlabeled events for self-supervised pretraining plus 70k labeled examples for track finding, PID, and noise tagging.","key_machinery":"The central object is the TPC spacepoint: a 4-entry record (E, x, y, z) of an ionization signal's position and amplitude left by a charged particle in a gas-filled time projection chamber. The unlabeled dataset contains only these arrays; the labeled dataset aligns them with truth track IDs, noise tags, and particle identities. Because every downstream task—pretraining, track finding, PID, noise tagging—is expressed as operations on the same spacepoint arrays, the spacepoint is the common representational unit that connects raw simulation to foundation-model training. The evaluation machinery is double-majority matching, which declares a predicted track matched to a true particle only when h","core_discovery":"The paper's central claim is that a 10-million-event simulated collision sample, stored as TPC spacepoints, can serve as the training and evaluation substrate for AI foundation models in charged-particle tracking. Each event is reduced to 4-column spacepoint records—ionization signal plus x, y, z position—plus, in the labeled subset, aligned track IDs, noise tags, and five-class particle-type labels. The authors assert the events are generated with a standard Monte Carlo event generator at 200 GeV, transported through a full detector simulation, digitized with channel-dependent gain and noise, and clustered into spacepoints, with the full chain reproducible from pinned software releases. For","pith_inferences":["An extension the paper leaves implicit: using only part of the 10M unlabeled events and measuring downstream task performance would directly probe scaling laws for self-supervised TPC representations.","The noise label is defined as spacepoints whose matched track has pT below 60 MeV/c; a benchmark built on this label may behave differently from one built on unmatched reconstructed fake spacepoints, so results should be read with this definition in mind.","A natural next experiment is testing whether models pretrained on these simulated spacepoints transfer to real TPC data; the release does not include a real-data comparison, so that transfer remains open."],"forward_implications":["Machine-learning groups can pretrain track-finding models on 10 million events without running particle-physics software or learning physics-specific data formats.","The 70,000 labeled events give the community a common benchmark: track finding, PID, and noise tagging can be compared across methods using the recommended metrics.","Because the full simulation and reconstruction chain is pinned to specific releases, the dataset can be regenerated or extended consistently when detector settings change.","TPC-specific effects—long drift distances, diffusion, and up to roughly 100 truth tracks per event—become testable for models trained on silicon-tracker data."],"supporting_citations":[{"why":"Supplies the Monte Carlo event-generation code used to produce the simulated proton–proton collisions at 200 GeV.","marker":"[1]"},{"why":"Supplies the underlying-event tune that fixes the event-generation settings for the simulated energy.","marker":"[2]"},{"why":"Supplies the particle-transport and detector-response simulation that turns generated events into detector signals.","marker":"[3]"},{"why":"Supplies the TPC geometry and spatial-resolution figures defining the detector configuration being simulated.","marker":"[4]"},{"why":"Supplies the particle-data codes that are grouped into the five PID categories.","marker":"[5]"},{"why":"Supplies the double-majority matching rule and the efficiency and purity definitions recommended for track-finding evaluation.","marker":"[6]"}],"fun_headline_variants":["10M simulated collisions for AI foundation models","Open dataset: 10M pp collisions for ML","70k labeled events for tracking, PID, noise tagging","10M simulated pp collisions for AI self-supervised training","Large simulated collision dataset for ML foundation models"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the simulated detector response—material, electronics, noise, and clustering—produces spacepoints that are a realistic stand-in for real TPC collisions; no comparison with real data is included.","fun_headline_variants_meta":{"raw":{"variants":["10M simulated collisions for AI foundation models","Open dataset: 10M pp collisions for ML","70k labeled events for tracking, PID, noise tagging","10M simulated pp collisions for AI self-supervised training","Large simulated collision dataset for ML foundation models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000936,"raw_usage":{"total_tokens":3872,"prompt_tokens":808,"completion_tokens":3064,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":2990}},"tokens_in":552,"tokens_out":3064,"duration_ms":23943,"temperature":1.0,"reasoning_tokens":2990,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T04:56:49.135312+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the spacepoint multiplicity, noise occupancy, and hit-position residuals in this dataset against real minimum-bias proton–proton data from the same TPC at 200 GeV; a large disagreement in any of these distributions would break the claim that the simulation is a faithful surrogate.","supporting_citations":[{"cited_title":"Klest, Overview and design of the sphenix tpc, Journal of Physics: Conference Series 1498 (1) (2020) 012025.doi:10.1088/1742-6596/1498/1/012025","cited_arxiv_id":null,"evidence_quote":"Supplies the TPC geometry and spatial-resolution figures defining the detector configuration being simulated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the double-majority matching rule and the efficiency and purity definitions recommended for track-finding evaluation."}],"review_version":1}