Pith. sign in

REVIEW 3 major objections 7 minor 79 references

Generalist Vision-Language Models for Fast Radio Burst detection: a zero-shot benchmark against a specialized detector

T0 review · 3 major / 7 minor · reviewed 2026-07-09 · glm-5.2

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Zero-shot AI matches specialized detector on FRB identification

desk verdict Solid first benchmark of zero-shot VLMs for FRB detection; the prompt is doing more work than the framing admits, but the experimental design is careful and the results are honestly reported. read the letter →

arxiv 2607.07382 v1 pith:C2WZS63R submitted 2026-07-08 cs.LG astro-ph.HEastro-ph.IM

Raiff H. Santos , Amilcar R. Queiroz , Tharcisyo S. S. Duarte , K. E. L. de Farias , Rafael A. Batista This is my paper · ORCID
classification cs.LGastro-ph.HEastro-ph.IM
keywords FastRadioBurstsVision-LanguageModelszero-shotclassificationfrequencyinterferencedynamicspectraSwinYNetGemmapromptengineering
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tests whether small generalist vision-language models (VLMs) can identify Fast Radio Bursts in radio telescope data without any task-specific training. The authors simulate 3,000 dynamic spectra—images showing signal power across time and frequency—containing FRBs, structured radio-frequency interference (RFI), and noise. They then compare two open-weight models (Gemma 4 2B and 4B), prompted only with a text description of what an FRB looks like, against SwinYNet, a specialized deep-learning detector trained specifically for FRB searches. The central finding is that Gemma 4 2B achieves 93.65% accuracy on the binary FRB-versus-non-FRB task, statistically indistinguishable from SwinYNet's 92.90% (McNemar p=0.33), while producing far fewer false positives on structured RFI (6.4% vs 25.0%, p=3.8×10⁻²¹) and zero false positives on pure noise. SwinYNet retains a perfect probabilistic ranking (ROC-AUC 1.0000 vs 0.9482), confirming it remains the ceiling. The VLM's exclusive errors are concentrated on the faintest FRBs, which it perceives as noise, while its exclusive gains are on RFI rejection—particularly multiband interference that fools the specialized detector. The paper also demonstrates that rewriting the prompt alone reconfigures the same model for three-class FRB/RFI/noise classification, reaching 86% accuracy with zero false FRBs.

What carries the argument

Vision-Language Models (VLMs) in zero-shot prompt-only regime; dynamic spectra as images; SwinYNet as specialized baseline; McNemar paired comparison; prompt-based task reconfiguration

What would settle it

Run the same zero-shot VLM protocol on real observational data from a telescope like CHIME or ASKAP, where RFI is uncontrolled and labels are uncertain. If the VLM's RFI rejection advantage disappears or its faint-FRB miss rate becomes unacceptable, the simulated-benchmark performance does not transfer.

Watch

Extended reading notes

Core claim

A generalist VLM with no FRB-specific training, operating only from a rendered image and a text prompt, approaches the discriminative performance of a dedicated detector on simulated data. The two systems exhibit complementary error profiles: SwinYNet catches all FRBs but misclassifies substantial RFI as bursts, while the VLM misses faint bursts but strongly rejects structured interference. The VLM's self-reported probability is heavily quantized (three distinct values for the 2B model), making its probabilistic metrics coarse. The prompt itself contains detailed morphological descriptions of FRBs and RFI types along with a probability calibration guide, meaning domain knowledge enters the系统

Load-bearing premise

The benchmark is entirely simulated with five controlled RFI categories and known ground-truth labels, and the prompt contains detailed domain-specific descriptions of FRB morphology and a probability calibration guide. If real RFI looks different from the simulated types, or if the prompt's domain knowledge is doing substantial work, the headline result may not transfer to real observations.

Editorial extensions

If this is right

  • Radio astronomy pipelines could deploy small local VLMs as candidate triage tools that reject RFI false positives more aggressively than specialized detectors, reducing the human inspection bottleneck.
  • The complementary error profiles suggest hybrid pipelines where a specialized detector ensures high FRB recall and a VLM filters its RFI false positives could outperform either alone.
  • Prompt-only task reconfiguration (binary to three-class without retraining) could be applied to other transient classification problems where labeled data is scarce.
  • The VLM's natural-language justifications for each decision, while not guaranteed physically correct, provide an audit trail absent from black-box detectors.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the prompt's domain-specific morphology descriptions and calibration guide are doing substantial work (as the detailed Appendix A suggests), then the 'zero-shot' label is partially misleading—this is prompt-engineered transfer, not pure visual recognition from general pretraining. Ablating the prompt's domain content would isolate how much performance comes from the model's latent visual knowle
  • The strong RFI rejection may not transfer to real data if observational RFI morphologies diverge from the five simulated categories, particularly if real interference exhibits dispersive-like sweeps or partial masking that the simulation does not capture.
  • The quantized probability output (three values for 2B) limits operational use as a ranking score; extracting continuous scores from token logits or using log-probability methods could substantially improve the VLM's utility as a probabilistic classifier.
  • The asymmetry where the VLM sees a PNG while SwinYNet sees native FITS data means the comparison conflates model capability with input representation—testing SwinYNet on rendered images or the VLM on richer numerical inputs would clarify how much of the gap is architectural versus representational.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. This manuscript evaluates whether small, open-weight generalist Vision-Language Models (VLMs) — Gemma 4 2B and 4B — can detect Fast Radio Bursts (FRBs) in dynamic spectra under a zero-shot, prompt-only regime, without any fine-tuning or labeled examples. The authors construct a controlled benchmark of 2000 simulated L-band dynamic spectra (1000 FRB, 500 RFI, 500 noise) and compare the VLMs sample-by-sample against the specialized detector SwinYNet. At the default 0.5 threshold, Gemma 4 2B achieves 93.65% accuracy with no statistically significant difference from SwinYNet (92.90%, McNemar p=0.3305), while showing a significantly lower false-positive rate on structured RFI (6.4% vs 25.0%, p=3.8e-21) and zero false positives on pure noise. SwinYNet retains a perfect probabilistic ranking (ROC-AUC 1.0000 vs 0.9482). The paper also demonstrates a three-class FRB/RFI/NOISE classification task via prompt rewriting, reaching up to 86% accuracy. The experimental design is careful: balanced benchmark, anonymization for leakage control, paired comparison, McNemar tests, bootstrap CIs, and honest reporting of all 2000 samples with no exclusions.

Significance. The paper addresses a timely question — whether generalist foundation models can approach specialized detectors on a scientific vision task — and provides a well-constructed benchmark with reproducible code and pinned simulation seeds. The paired sample-by-sample comparison with McNemar tests and bootstrap CIs is methodologically sound. The stratified false-positive analysis (RFI vs noise) and the characterization of disagreement cases (VLM excels on structured RFI, fails on faint FRBs) are genuinely informative. The multiclass demonstration via prompt rewriting alone, with zero false FRBs, is a clean illustration of the flexibility claim. The honest reporting of score quantization (three values for Gemma 4 2B) and the threshold-artifact caveat for the accuracy parity are commendable. The central empirical finding — that a zero-shot VLM approaches a specialized detector at the default threshold while exhibiting a complementary error profile — is interesting and the evidence is presented fairly.

major comments (3)
  1. §III B and Abstract: The paper defines zero-shot as relying 'solely on the representations acquired during general multimodal pretraining and on a textual instruction that specifies the candidate classes' (§III B). However, the binary prompt (Appendix A) goes well beyond specifying candidate classes. It encodes: (1) a detailed FRB morphology definition ('SINGLE localized transient that is plausibly broadband and shows a frequency-dependent arrival-time trend, usually later arrival at lower frequency; the sweep may be curved, nearly linear, faint, or partially masked'); (2) explicit NON_FRB decision criteria enumerating specific RFI morphologies (persistent bands, zero-DM impulses, periodic stripes, etc.); (3) a probability calibration guide with 8 mapped ranges; (4) low-S/N sensitivity instructions; and (5) definitions for 9 structured visual features. This is a detailed expert decision,
  2. §VI D and Table IX: The most striking result — the VLM's significantly lower false-positive rate on structured RFI (6.4% vs 25.0%, p=3.8e-21) — may depend substantially on the prompt's explicit enumeration of RFI morphologies to reject (Appendix A, rule 3: 'persistent horizontal frequency bands, persistent vertical time bands, periodic stripes or grid-like structure, zero-DM broadband vertical impulse'). The multiclass prompt (Appendix B), which deliberately avoids enumerating RFI morphologies, shows much lower RFI recall (0.534 for 2B, 0.660 for 4B), suggesting that the binary prompt's explicit RFI descriptions are doing substantial work. Without a prompt ablation (e.g., a minimal prompt that only specifies the two class labels without morphology descriptions), the attribution of RFI rejection to 'general-purpose pretraining alone' is untested. This is load-bearing for the central claim
  3. §VI A and §VI B: The 'no statistically significant difference' claim (McNemar p=0.3305) is presented in the abstract and conclusions as a headline result, but the paper itself acknowledges (§VI B) that SwinYNet achieves perfect probabilistic ranking (ROC-AUC 1.0000) while the VLM achieves 0.9482, and that the accuracy parity at 0.5 is partly a threshold artifact. The bootstrap CI for the ROC-AUC delta is [-0.0616, -0.0423], clearly excluding zero. The paper handles this tension honestly in the body, but the abstract and conclusions ('no statistically significant difference from SwinYNet') could mislead readers who do not read the probabilistic analysis. The framing should be adjusted so that the abstract does not imply parity in discriminative ability, only in discrete decisions at one threshold.
minor comments (7)
  1. §V B: The models are referred to as 'Gemma 4 2B' and 'Gemma 4 4B' throughout, citing Ref. [39] as 'Gemma 4: Frontier Multimodal Intelligence on Device' (2026). As of the manuscript's stated date, this model family may not be widely verifiable. The authors should confirm the exact model version and checkpoint used, and ideally pin the Hugging Face model identifier for full reproducibility.
  2. Table V: The column header 'Precision (NON FRB)' and 'Recall (NON FRB)' could be more precisely labeled as 'Precision (NON_FRB)' for consistency with the label notation used elsewhere.
  3. §IV D: The flux density range is given as 'approximately 4.02 to 11.98 in the internal units used by the simulation.' It would help readers if the authors clarified what these internal units represent (e.g., Jy or arbitrary simulation units), since the faint-FRB analysis in §VI F references percentiles of this quantity.
  4. Fig. 2: The three representative images are useful but small. Consider adding annotations (e.g., arrows pointing to the dispersed sweep in the FRB panel) to help non-specialist readers identify the visual features being discussed.
  5. §VI G, Table X: The per-class metrics for the multiclass task show perfect NOISE recall (1.0000) for both models but substantially lower RFI recall. The text explains this well, but it would strengthen the presentation to include a brief confusion matrix for the multiclass task as well, not just per-class metrics.
  6. Appendix A: The prompt instruction 'Use the full range and avoid always reusing the same values' appears to be an attempt to address score quantization, but §VI B reports that Gemma 4 2B used only three values. This is worth a brief comment in the prompt design discussion — the instruction was present but ineffective for the smaller model.
  7. §II C: The complexity of incoherent dedispersion is stated as O(N_t * N_nu * N_DM), but the symbol N_nu is not defined in the surrounding text (N_t and N_DM are defined). Please add the definition.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive report. The referee raises three major comments concerning (1) the definition of 'zero-shot' given the richness of the binary prompt, (2) the absence of a prompt ablation to test whether RFI rejection is attributable to pretraining versus the prompt's explicit RFI morphology enumeration, and (3) the framing of the McNemar parity result in the abstract and conclusions relative to the probabilistic ranking gap. We agree with all three points and will revise the manuscript accordingly. Specifically, we will (a) correct the zero-shot definition to distinguish between 'no labeled examples / no weight updates' and 'no task-specific textual guidance,' (b) add a minimal-prompt ablation to the revised manuscript, and (c) adjust the abstract and conclusions to restrict the parity claim to discrete decisions at the 0.5 threshold rather than implying overall discriminative parity. We believe these revisions strengthen the paper without altering its core contribution.

read point-by-point responses

  1. Referee: The paper defines zero-shot as relying 'solely on the representations acquired during general multimodal pretraining and on a textual instruction that specifies the candidate classes' (§III B). However, the binary prompt (Appendix A) goes well beyond specifying candidate classes. It encodes: (1) a detailed FRB morphology definition; (2) explicit NON_FRB decision criteria enumerating specific RFI morphologies; (3) a probability calibration guide with 8 mapped ranges; (4) low-S/N sensitivity instructions; and (5) definitions for 9 structured visual features. This is a detailed expert decision procedure, not merely a class specification.

    Authors: The referee is correct. The definition in §III B is too narrow: it states that the textual instruction 'specifies the candidate classes,' but the binary prompt in Appendix A goes substantially further, encoding expert morphological descriptions, explicit RFI rejection criteria, a calibration guide, and feature definitions. This is a detailed decision procedure, not a bare class label specification. We will revise the definition in §III B to distinguish clearly between two properties that the current text conflates: (1) the zero-shot property in the standard machine-learning sense — no labeled examples shown, no weight updates — which holds; and (2) the claim that the prompt 'specifies only the candidate classes,' which does not hold for the binary prompt. We will rewrite the definition to state that the models operate without labeled examples or weight updates, but that the binary prompt provides expert morphological guidance including FRB and RFI descriptions, a calibration guide, and feature definitions. We will also add an explicit acknowledgment that the prompt encodes domain knowledge and that the performance therefore reflects what the VLM can do when given expert textual guidance, not what it can do from class labels alone. The multiclass prompt (Appendix B), which is deliberately minimal and defines RFI as a residual class without enumerating morphologies, will be highlighted as a contrasting point of reference. This revision does not undermine the paper's core claim — the models still receive no labeled examples and no weight updates — but it makes the boundary between 'zero-shot' and 'prompt-engineered' honest. revision: yes

  2. Referee: The most striking result — the VLM's significantly lower false-positive rate on structured RFI (6.4% vs 25.0%) — may depend substantially on the prompt's explicit enumeration of RFI morphologies to reject. The multiclass prompt, which deliberately avoids enumerating RFI morphologies, shows much lower RFI recall, suggesting that the binary prompt's explicit RFI descriptions are doing substantial work. Without a prompt ablation (e.g., a minimal prompt that only specifies the two class labels without morphology descriptions), the attribution of RFI rejection to 'general-purpose pretraining alone' is untested.

    Authors: This is a fair and important point, and we agree that the attribution of RFI rejection to 'general-purpose pretraining alone' is untested without a prompt ablation. The referee's observation about the multiclass prompt is insightful: the multiclass prompt defines RFI as a residual class without enumerating morphologies, and indeed RFI recall drops substantially (0.534 for 2B, 0.660 for 4B), which is consistent with the hypothesis that the binary prompt's explicit RFI descriptions contribute to the strong RFI rejection. We will run a minimal-prompt ablation: a binary prompt that specifies only the two class labels (FRB and NON_FRB) with a brief definition of each, without enumerating RFI morphologies, without the calibration guide, and without the feature definitions block. We will report the resulting accuracy, RFI false-positive rate, and noise false-positive rate alongside the current binary prompt results. This will allow us to decompose the contribution of the expert prompt from the contribution of pretrained visual representations. We expect, based on the multiclass evidence the referee cites, that the RFI rejection will weaken with the minimal prompt, and we will report whatever the result is honestly. We will also revise all instances in the manuscript where RFI rejection is attributed to 'general-purpose pretraining alone' to instead say that it reflects the combination of pretrained representations and the expert prompt, pending the ablation results to quantify each factor's contribution. revision: yes

  3. Referee: The 'no statistically significant difference' claim (McNemar p=0.3305) is presented in the abstract and conclusions as a headline result, but the paper itself acknowledges that SwinYNet achieves perfect probabilistic ranking (ROC-AUC 1.0000) while the VLM achieves 0.9482, and that the accuracy parity at 0.5 is partly a threshold artifact. The bootstrap CI for the ROC-AUC delta is [-0.0616, -0.0423], clearly excluding zero. The framing should be adjusted so that the abstract does not imply parity in discriminative ability, only in discrete decisions at one threshold.

    Authors: The referee is correct. The abstract and conclusions currently state 'no statistically significant difference from SwinYNet' without immediately qualifying that this refers only to discrete decisions at the 0.5 threshold, and that SwinYNet is strictly superior as a probabilistic classifier (ROC-AUC 1.0000 vs 0.9482, bootstrap CI excluding zero). The body of the paper handles this tension honestly in §VI B, but the framing in the abstract and conclusions could mislead a reader who does not read that section. We will revise the abstract to state that Gemma 4 2B shows no statistically significant difference from SwinYNet in discrete classification at the default 0.5 threshold (McNemar p=0.3305), while making clear in the same sentence or the next that SwinYNet retains a perfect probabilistic ranking (ROC-AUC 1.0000 vs 0.9482). We will apply the same revision to the conclusions in §VII. The current abstract already contains the ROC-AUC sentence, but it follows the parity claim without explicitly characterizing the parity as threshold-specific; we will reorder and qualify so that the threshold-specificity of the parity claim is clear before the ROC-AUC is introduced. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical benchmark with external baseline, no derivation chain to reduce

full rationale

This paper is an empirical benchmark, not a derivation. There is no equation chain where outputs are defined in terms of inputs by construction. The VLM (Gemma 4) is a pretrained model with frozen weights from external pretraining; SwinYNet is an independently trained external detector used as-is from its official repository. The comparison between them is paired on the same samples but neither system's predictions are defined by the other's. The prompt (Appendix A) encodes domain knowledge, which is a validity concern (acknowledged in §VI H), but this is a question of attribution and external validity, not circularity. The metrics (accuracy, ROC-AUC, McNemar test) are standard and computed independently of both models' internal mechanisms. No self-citation chain is load-bearing for the central claim: the paper cites SwinYNet [31] as an external baseline, simulateSearch [44] as an external simulation tool, and the Gemma models [39] as external pretrained weights. The authors' prior cosmology papers [14-20] are referenced for background motivation and do not form a premise-prediction loop. The 'zero-shot from general-purpose pretraining alone' framing is a claim about what the VLM has seen, which is independently falsifiable (e.g., by testing on novel RFI morphologies or with prompt ablations). The absence of a prompt ablation is a gap in experimental design, not a circularity: the prompt's domain knowledge is not used to define the evaluation metric or the ground-truth labels. The derivation is self-contained against external benchmarks.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities, particles, forces, or dimensions are introduced.

free parameters (3)
  • Decision threshold (0.5) = 0.5
    The binary decision threshold is fixed at 0.5 for all models. This is not fitted to data but is a design choice that affects the accuracy comparison. SwinYNet's 142 false positives at this threshold vs. its perfect ROC-AUC show the threshold is suboptimal for the specialized detector.
  • Image rendering parameters = viridis colormap, percentile normalization (1st-99th), 1024×768 px
    These are fixed design choices that affect VLM performance. The paper acknowledges (§VI H, point 4) that colormap, normalization, resolution, and contrast can affect the decision, but no ablation is performed.
  • System temperature scale range = 0.650187 to 1.449952 (dimensionless)
    The per-sample noise scale range is a simulation parameter chosen to prevent global brightness from being a class cue. It is not fitted to model performance but constrains the difficulty distribution.
assumptions (4)
  • domain assumption Simulated dynamic spectra with controlled RFI categories are representative enough of real observations to support claims about detector transferability.
    Invoked throughout §IV and acknowledged as a limitation in §VI H (point 1). The entire benchmark and all results depend on this assumption.
  • ad hoc to paper A text prompt describing FRB morphology in detail does not constitute 'training' or 'supervision' for the purpose of the zero-shot claim.
    The prompt in Appendix A contains detailed decision rules, a probability calibration guide with specific numeric ranges, and feature definitions. The paper frames this as zero-shot (§III B, §V), but the prompt encodes substantial domain expertise. This axiom is load-bearing for the 'zero-shot' framing.
  • domain assumption General-purpose visual features learned during multimodal pretraining transfer to scientific dynamic-spectrum images.
    Invoked in §III C and §III D. The paper's thesis depends on pretrained features being expressive enough to separate dispersed bursts from interference without domain-specific training.
  • ad hoc to paper The VLM's self-reported frb_probability is a valid continuous score for ROC-AUC and calibration analysis.
    §V C states the probability is 'a number that the model chooses' rather than a quantity derived from token probabilities. The paper acknowledges (§VI B) that this score is heavily quantized (3-9 distinct values), but still uses it for ROC-AUC, AP, Brier, and ECE computation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generalist Vision-Language Models for Fast Radio Burst detection: a zero-shot benchmark against a specialized detector." pith.science (2026). https://pith.science/paper/C2WZS63R

@misc{pith2026260707382,
  author       = {Pith},
  title        = {Pith review of: Generalist Vision-Language Models for Fast Radio Burst detection: a zero-shot benchmark against a specialized detector},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C2WZS63R}},
  note         = {Machine review of arXiv:2607.07382}
}
read the original abstract

Fast Radio Bursts (FRBs) are millisecond-duration radio transients whose automated detection increasingly relies on highly specialized deep learning models. These detectors achieve exceptional performance, but they require large task-specific training datasets and cannot be redefined without retraining. In this work, we evaluate whether small, open-weight, locally run generalist Vision-Language Models (VLMs) can detect FRBs in dynamic spectra under a zero-shot, prompt-only regime, with no fine-tuning and no labeled examples, returning structured decisions with a natural-language justification. From a controlled set of 3000 simulated L-band dynamic spectra containing FRBs, structured Radio Frequency Interference (RFI), and noise, we draw a balanced binary benchmark of 2000 samples and compare two such VLMs (Gemma 4 2B and 4B), sample by sample, against the state-of-the-art specialized detector SwinYNet. At the default threshold, Gemma 4 2B reaches an accuracy of 93.65%, with no statistically significant difference from SwinYNet (92.90%), while showing a significantly lower false-positive rate on structured RFI (6.4% vs. 25.0%) and no false positives on pure noise. SwinYNet retains a perfect probabilistic ranking on this benchmark (ROC-AUC of 1.0000 vs. 0.9482), a ceiling that the zero-shot VLM approaches from general-purpose pretraining alone. Rewriting the prompt alone reconfigures the same models for three-class FRB/RFI/noise classification on the full set of 3000 spectra, where they reach up to 86% accuracy without a single false FRB.

Figures

Figures reproduced from arXiv: 2607.07382 by the authors.

Figure 1
Figure 1. FIG. 1. Overview of the end-to-end pipeline, from simulation to results. A controlled set of 3000 dynamic spectra (1000 [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Representative anonymized dynamic-spectrum images, one per class, as supplied to the VLM. Left: an [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. ROC curves for the binary [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: FIG. 4. Histograms of the predicted [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

79 extracted references · 79 canonical work pages

  1. [1]

    The sweep may be curved, nearly linear, faint, or partially masked

    FRB candidates Label as FRB only when the dominant signal is a SINGLE localized transient that is plausibly broadband and shows a frequency-dependent arrival-time trend, usually later arrival at lower frequency. The sweep may be curved, nearly linear, faint, or partially masked

    work page

  2. [2]

    Low-S/N sensitivity Do not treat faint coherent structure as pure noise. If there is any plausible faint diagonal/curved sweep, localized broadband brightening, partial dispersed track, or excess variance along a tilted band, use an intermediate probability instead of a very low one

    work page

  3. [3]

    NON_FRB cases Label as NON_FRB when the image is dominated by: - pure random background with no coherent transient, - persistent horizontal frequency bands, - persistent vertical time bands, - periodic stripes or grid-like structure, - zero-DM broadband vertical impulse with no dispersion delay, - saturation, clipping, blank regions, uniform panels, or no...

    work page

  4. [4]

    FRB" only if frb_probability >= 0.50 - label =

    Probability calibration frb_probability must be in [0,1]. 20 Use the full range and avoid always reusing the same values. Calibration guide: - 0.95-0.99: textbook FRB, clear broadband dispersed sweep. - 0.80-0.94: probable FRB with minor ambiguity. - 0.60-0.79: plausible FRB but faint, partial, or noisy. - 0.45-0.59: genuinely ambiguous. - 0.30-0.44: weak...

    work page

  5. [5]

    Is there a single broadband sweep that drifts monotonically in frequency with time? If yes, label FRB

    work page

  6. [6]

    Otherwise, is the panel pure random texture with no coherent feature at all? If yes, label NOISE

    work page

  7. [7]

    label":

    Otherwise, label RFI. Rules ===== - NOISE requires a complete absence of coherent structure. If you can see ANY non-random feature anywhere in the panel -- even faint, even sparse, even a single thin line, a streak, or a few repeated marks -- the image is NOT NOISE. - Faint or sparse structure still counts as structure. - An FRB drifts in frequency and ap...

    work page

  8. [8]

    A bright millisecond radio burst of extragalactic origin,

    D. R. Lorimer, M. Bailes, M. A. McLaughlin, D. J. Narkevic, and F. Crawford, “A bright millisecond radio burst of extragalactic origin,”Science, vol. 318, no. 5851, pp. 777–780, 2007

    work page 2007

Show all 79 references
  1. [9]

    Fast radio bursts—a brief review: Some ques- tions, fewer answers,

    J. I. Katz, “Fast radio bursts—a brief review: Some ques- tions, fewer answers,”Modern Physics Letters A, vol. 31, no. 14, p. 1630013, 2016

    2016

  2. [10]

    Fast radio bursts,

    E. Petroff, J. W. T. Hessels, and D. R. Lorimer, “Fast radio bursts,”The Astronomy and Astrophysics Review, vol. 27, no. 1, p. 4, 2019

    2019

  3. [11]

    The physical mechanisms of fast radio bursts,

    B. Zhang, “The physical mechanisms of fast radio bursts,”Nature, vol. 587, pp. 45–53, 2020

    2020

  4. [12]

    The physics of fast radio bursts,

    B. Zhang, “The physics of fast radio bursts,”Reviews of Modern Physics, vol. 95, no. 3, p. 035005, 2023

    2023

  5. [13]

    A needle in a cos- mic haystack: A review of FRB search techniques,

    K. M. Rajwade and J. van Leeuwen, “A needle in a cos- mic haystack: A review of FRB search techniques,”Uni- verse, vol. 10, no. 4, p. 158, 2024

    2024

  6. [14]

    The properties of fast ra- dio bursts: A short review,

    W.-Q. Ma and Z.-F. Gao, “The properties of fast ra- dio bursts: A short review,”Astronomische Nachrichten, vol. 346, no. 7, p. e20250024, 2025

    2025

  7. [15]

    A real-time fast radio burst: polar- ization detection and multiwavelength follow-up,

    E. Petroffet al., “A real-time fast radio burst: polar- ization detection and multiwavelength follow-up,”Mon. Not. Roy. Astron. Soc., vol. 447, no. 1, pp. 246–255, 2015

    2015

  8. [16]

    Fast Radio Bursts: An Extragalactic Enigma,

    J. M. Cordes and S. Chatterjee, “Fast Radio Bursts: An Extragalactic Enigma,”Ann. Rev. Astron. Astrophys., vol. 57, pp. 417–465, 2019

    2019

  9. [17]

    Canadian hydrogen in- tensity mapping experiment (CHIME) pathfinder,

    K. Bandura, G. E. Addison, M. Amiri, J. R. Bond, D. Campbell-Wilson, L. Connor, J.-F. Cliche, G. Davis, M. Deng, N. Denman,et al., “Canadian hydrogen in- tensity mapping experiment (CHIME) pathfinder,” in Ground-based and Airborne Telescopes V, vol. 9145, p. 914522, SPIE, 2014

    2014

  10. [18]

    Science with the Australian square kilometre array pathfinder,

    S. Johnston, M. Bailes, N. Bartel, C. Baugh, M. Bieten- holz, C. Blake, R. Braun, J. Brown, S. Chatterjee, J. Dar- ling,et al., “Science with the Australian square kilometre array pathfinder,”Publications of the Astronomical So- ciety of Australia, vol. 24, no. 4, pp. 174–188, 2007

    2007

  11. [19]

    The five-hundred-meter aperture spherical radio telescope (FAST) project,

    R. Nan, D. Li, C. Jin, Q. Wang, L. Zhu, W. Zhu, H. Zhang, Y. Yue, and L. Qian, “The five-hundred-meter aperture spherical radio telescope (FAST) project,”In- ternational Journal of Modern Physics D, vol. 20, no. 06, pp. 989–1024, 2011

    2011

  12. [20]

    Revival of the magnetar PSR J1622–4950: Observations with MeerKAT, Parkes, XMM-Newton, Swift, Chandra, and NuSTAR,

    F. Camilo, P. Scholz, M. Serylak, S. Buchner, M. Mer- ryfield, V. Kaspi, R. Archibald, M. Bailes, A. Jame- son, W. v. Straten,et al., “Revival of the magnetar PSR J1622–4950: Observations with MeerKAT, Parkes, XMM-Newton, Swift, Chandra, and NuSTAR,”The As- trophysical Journal...

    2018

  13. [21]

    Cosmological model-independent constraints on the baryon fraction in the IGM from fast radio bursts and supernovae data,

    T. Lemos, R. S. Gon¸ calves, J. C. Carvalho, and J. S. Al- caniz, “Cosmological model-independent constraints on the baryon fraction in the IGM from fast radio bursts and supernovae data,”Eur. Phys. J. C, vol. 83, no. 2, p. 138, 2023

    2023

  14. [22]

    Cosmography from well-localized fast radio bursts,

    J. A. S. Fortunato, W. S. Hip´ olito-Ricaldi, and M. V. 22 dos Santos, “Cosmography from well-localized fast radio bursts,”Mon. Not. Roy. Astron. Soc., vol. 526, no. 2, pp. 1773–1782, 2023

    2023

  15. [23]

    Cosmological model-independent limits on photon mass from FRB and SNe data,

    T. Lemos, R. Gon¸ calves, J. Carvalho, and J. Alcaniz, “Cosmological model-independent limits on photon mass from FRB and SNe data,”JCAP, vol. 11, p. 019, 2025

    2025

  16. [24]

    Cosmo- graphic constraints from late-time probes including fast radio bursts,

    L. L. Sales, K. E. L. de Farias, A. R. Queiroz, J. R. L. Santos, R. A. Batista, A. R. M. Oliveira, L. F. San- tana, C. A. Wuensche, T. Villela, and J. Vieira, “Cosmo- graphic constraints from late-time probes including fast radio bursts,” 7 2025

    2025

  17. [25]

    Constraints on the baryon density from fast radio bursts using a non- parametric reconstruction of the Hubble parameter,

    L. L. Sales, K. E. L. de Farias, A. R. Queiroz, R. A. Batista, B. W. Ribeiro, and R. H. Santos, “Constraints on the baryon density from fast radio bursts using a non- parametric reconstruction of the Hubble parameter,” 5 2026

    2026

  18. [26]

    Prob- ing Primordial Black Holes with upcoming Radio Tele- scopes: a case study for LOFAR2.0, FAST Core Array and BINGO,

    J. R. L. Santos, G. Dom` enech, and A. R. Queiroz, “Prob- ing Primordial Black Holes with upcoming Radio Tele- scopes: a case study for LOFAR2.0, FAST Core Array and BINGO,” 4 2026

    2026

  19. [27]

    Constraints on Dark Energy and Modified Gravity Models from Fast Radio Bursts and Late-Time Geometric Probes,

    B. W. N. Ribeiro, L. L. Sales, K. E. L. de Farias, R. H. Santos, R. A. Batista, and A. R. Queiroz, “Constraints on Dark Energy and Modified Gravity Models from Fast Radio Bursts and Late-Time Geometric Probes,” 6 2026

    2026

  20. [28]

    BASSET: Bandpass-adaptive single-pulse search toolkit—optimized subband pulse search strategies for faint narrowband fast radio bursts,

    J. H. Cao, P. Wang, D. Li, Q. H. Pan, K. Mao, C. H. Niu, Y. K. Zhang, Q. Y. Qu, W. J. Lu, and J. S. Zhang, “BASSET: Bandpass-adaptive single-pulse search toolkit—optimized subband pulse search strategies for faint narrowband fast radio bursts,”The Astrophysical Journal Supplem...

    2025

  21. [29]

    Searching for short-time-scale radio anomalies using non-linear di- mensionality reduction techniques,

    X. Yang, G. Hobbs, S.-B. Zhang, A. Zic, L. Toomey, Y. Li, J.-S. Wang, S. Dai, and X.-F. Wu, “Searching for short-time-scale radio anomalies using non-linear di- mensionality reduction techniques,”Monthly Notices of the Royal Astronomical Society, vol. 537, pp. 991–1000, February 2025

    2025

  22. [30]

    Clas- sifying frb spectrograms using nonlinear dimensionality reduction techniques,

    X. Yang, S.-B. Zhang, J.-S. Wang, and X.-F. Wu, “Clas- sifying frb spectrograms using nonlinear dimensionality reduction techniques,”Monthly Notices of the Royal As- tronomical Society, vol. 522, pp. 4342–4351, July 2023

    2023

  23. [31]

    S. M. Ransom,New search techniques for binary pulsars. PhD thesis, Harvard University, Cambridge, MA, 2001. ADS Bibcode: 2001PhDT.......123R

    2001

  24. [32]

    PRESTO: PulsaR exploration and search TOolkit

    S. M. Ransom, “PRESTO: PulsaR exploration and search TOolkit.” Astrophysics Source Code Library, record ascl:1107.017, 2011.https://ascl.net/1107. 017

    2011

  25. [33]

    Accelerating inco- herent dedispersion,

    B. R. Barsdell, M. Bailes, D. G. Barnes, C. J. Fluke, A. Jameson, and I. G. Bebbington, “Accelerating inco- herent dedispersion,”Monthly Notices of the Royal As- tronomical Society, vol. 422, pp. 379–392, May 2012

    2012

  26. [34]

    D. R. Lorimer and M. Kramer,Handbook of pulsar as- tronomy, vol. 4. Cambridge university press, 2005

    2005

  27. [35]

    Enhancing fast radio transient detection with mask r-cnn image segmentation,

    S. Belmonte D´ ıazet al., “Enhancing fast radio transient detection with mask r-cnn image segmentation,”RAS Techniques and Instruments, vol. 5, no. 1, p. rzag006, 2026

    2026

  28. [36]

    Accelerating the fast radio burst search: Data set and methods,

    X. Guo, H. Wang, Y. Xiao, H. Chen, Y. Ke, C. Miao, P. Wang, D. Li, C. Jin, L. He,et al., “Accelerating the fast radio burst search: Data set and methods,”The As- trophysical Journal Supplement Series, vol. 280, no. 1, p. 34, 2025

    2025

  29. [37]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” inInternational Conference on Learning ...

    2021

  30. [38]

    SwinYNet: A transformer-based multitask model for accurate and effi- cient fast radio burst searches,

    Y. Chen, S. Ni, C. Li, J. Fang, D. Zhou, H. Chen, Y. Feng, P. Wang, C. Jin, H. Wang,et al., “SwinYNet: A transformer-based multitask model for accurate and effi- cient fast radio burst searches,”The Astrophysical Jour- nal Supplement Series, vol. 283, no. 2, p. 45, 2026

    2026

  31. [39]

    A survey on multimodal large language models,

    S. Yin, C. Fu, S. Zhao, K. Li, X. Sun, T. Xu, and E. Chen, “A survey on multimodal large language models,”Na- tional Science Review, vol. 11, p. nwae403, 12 2024

    2024

  32. [40]

    AstroSight: Galaxy morphology classification with multimodal large language models,

    B. He, F. Zhao, Y. Wang, P. Chen, X. Wang, D. Ma, H. Xu, and H. Li, “AstroSight: Galaxy morphology classification with multimodal large language models,” Publications of the Astronomical Society of the Pacific, vol. 138, no. 4, p. 044508, 2026

    2026

  33. [41]

    Do vision-language models see dwarf galax- ies the way we do?,

    D. Tanoglidis, C. Y. Tan, K. Overdeck, and A. Drlica- Wagner, “Do vision-language models see dwarf galax- ies the way we do?,”arXiv preprint arXiv:2606.07779,

    arXiv

  34. [42]

    Accepted at the Conference on Physics and AI (PAI 2026)

    2026

  35. [43]

    AstroLLaVA: to- wards the unification of astronomical data and natural language,

    S. Zaman, M. J. Smith, P. Khetarpal, R. Chakrabarty, M. Ginolfi, M. Huertas-Company, M. Jab lo´ nska, S. Kruk, M. L. Lain, S. J. R. M´ endez,et al., “AstroLLaVA: to- wards the unification of astronomical data and natural language,”arXiv preprint arXiv:2504.08583, 2025. Ac- cep...

    2025 arXiv

  36. [44]

    Talking with the latents — how to con- vert your LLM into an astronomer,

    I. Kamai, M. Huertas-Company, M. J. Smith, and H. B. Perets, “Talking with the latents — how to con- vert your LLM into an astronomer,”arXiv preprint arXiv:2602.09670, 2026

    2026

  37. [45]

    Textual inter- pretation of transient image classifications from large language models,

    F. Stoppa, T. Bulmus, S. Bloemen, S. J. Smartt, P. J. Groot, P. Vreeswijk, and K. W. Smith, “Textual inter- pretation of transient image classifications from large language models,”Nature Astronomy, vol. 9, no. 12, pp. 1869–1878, 2025

    2025

  38. [46]

    Gemini: a family of highly capable multi- modal models,

    Gemini Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Mil- lican,et al., “Gemini: a family of highly capable multi- modal models,”arXiv preprint arXiv:2312.11805, 2023

    2023 arXiv

  39. [47]

    Gemma 4: Frontier Multimodal Intel- ligence on Device

    Gemma Team, “Gemma 4: Frontier Multimodal Intel- ligence on Device.”https://deepmind.google/models/ gemma/gemma-4/, 2026. Released April 2026. Open weights under Apache 2.0. Google DeepMind

    2026

  40. [48]

    Pulsar candidate clas- sification with multimodal large language models,

    F. Zhao, Y. Li, Y. Wang, H. Li, M. Chen, P. Chen, N. Sun, C. Wang, and J. Liu, “Pulsar candidate clas- sification with multimodal large language models,” in NeurIPS 2024 Workshop on Foundation Models for Sci- ence: Progress, Opportunities, and Challenges, 2024

    2024

  41. [49]

    Unveiling the power of multimodal large language models for radio astronomical image under- standing and question answering,

    F. Zhao, Y. Li, Z. Liu, P. Chen, C. Wang, J. Liu, H. Li, and Y. Wang, “Unveiling the power of multimodal large language models for radio astronomical image under- standing and question answering,”Machine Learning: Science and Technology, vol. 6, no. 4, p. 045005, 2025

    2025

  42. [50]

    radio-llava: Ad- vancing vision-language models for radio astronomical source analysis,

    S. Riggi, T. Cecconello, A. Pilzer, S. Palazzo, N. Gupta, A. Hopkins, C. Trigilio, and G. Umana, “radio-llava: Ad- vancing vision-language models for radio astronomical source analysis,”Publications of the Astronomical Soci- ety of Australia, vol. 42, p. e121, 2025

    2025

  43. [51]

    Radio astronomy in the era of vision-language models: Prompt sensitivity and adaptation,

    M. Drozdova, E. Lastufka, V. Kinakh, T. Holotyak, D. Schaerer, and S. Voloshynovskiy, “Radio astronomy in the era of vision-language models: Prompt sensitivity and adaptation,”arXiv preprint arXiv:2509.02615, 2025. 23 Presented at the Machine Learning and the Physical Sci- enc...

    2025

  44. [52]

    Simulating high-time resolu- tion radio-telescope observations,

    R. Luo, G. Hobbs, S. Y. Yong, A. Zic, L. Toomey, S. Dai, A. Dunning, D. Li, T. Marshman, C. Wang, P. Wang, S. Wang, and S. Zhang, “Simulating high-time resolu- tion radio-telescope observations,”Monthly Notices of the Royal Astronomical Society, vol. 513, no. 4, pp. 5881– 5891, 2022

    2022

  45. [53]

    A fast ra- dio burst associated with a galactic magnetar,

    C. D. Bochenek, V. Ravi, K. V. Belov, G. Hallinan, J. Kocz, S. R. Kulkarni, and D. L. McKenna, “A fast ra- dio burst associated with a galactic magnetar,”Nature, vol. 587, pp. 59–62, 2020

    2020

  46. [54]

    A census of baryons in the uni- verse from localized fast radio bursts,

    J.-P. Macquart, J. Prochaska, M. McQuinn, K. Bannis- ter, S. Bhandari, C. Day, A. Deller, R. Ekers, C. James, L. Marnoch,et al., “A census of baryons in the uni- verse from localized fast radio bursts,”Nature, vol. 581, no. 7809, pp. 391–395, 2020

    2020

  47. [55]

    A new electron- density model for estimation of pulsar and frb distances,

    J. Yao, R. Manchester, and N. Wang, “A new electron- density model for estimation of pulsar and frb distances,” The Astrophysical Journal, vol. 835, no. 1, p. 29, 2017

    2017

  48. [56]

    NE2001. i. a new model for the galactic distribution of free electrons and its fluctuations,

    J. M. Cordes and T. J. W. Lazio, “NE2001. i. a new model for the galactic distribution of free electrons and its fluctuations,”arXiv preprint astro-ph/0207156, 2002

    2002 arXiv

  49. [57]

    DRAFTS: A deep- learning-based radio fast transient search pipeline,

    Y.-K. Zhang, D. Li, Y. Feng, C.-W. Tsai, P. Wang, C.- H. Niu, H.-X. Chen, and Y.-H. Zhu, “DRAFTS: A deep- learning-based radio fast transient search pipeline,”The Astrophysical Journal Supplement Series, vol. 276, no. 1, p. 20, 2025

    2025

  50. [58]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y. J. Lee, “Visual instruction tuning,” inAdvances in Neural Information Processing Systems(A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, eds.), vol. 36, pp. 34892–34916, Curran Associates, Inc., 2023

    2023

  51. [59]

    A sys- tematic evaluation of vision-language models for obser- vational astronomical reasoning tasks,

    W. Ren, H. Guo, W. Zuo, and X. Zhang, “A sys- tematic evaluation of vision-language models for obser- vational astronomical reasoning tasks,”arXiv preprint arXiv:2604.24589, 2026

    2026 arXiv

  52. [60]

    AstroAlertBench: Evaluating the accuracy, rea- soning, and honesty of multimodal LLMs in astronomical classification,

    C. Chen, J. S. Xiao, S. D. Liu, F. P. Paolino, L. Handley, T. J. d. Laz, R. Nilsson, A. Zou, M. Graham, and A. Ma- habal, “AstroAlertBench: Evaluating the accuracy, rea- soning, and honesty of multimodal LLMs in astronomical classification,”arXiv preprint arXiv:2605.05573, 2026

    2026 arXiv

  53. [61]

    AstroMMBench: A benchmark for evalu- ating multimodal large language models capabilities in astronomy,

    J. Shi, X. Tang, Y. Huang, Y. Li, X. Kong, Y. Zhang, and C. Yue, “AstroMMBench: A benchmark for evalu- ating multimodal large language models capabilities in astronomy,”arXiv preprint arXiv:2510.00063, 2025

    2025

  54. [62]

    Learning transferable vi- sual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable vi- sual models from natural language supervision,” inPro- ceedings of the 38th International Conference on Machin...

    2021

  55. [63]

    Swin transformer: Hierarchical vi- sion transformer using shifted windows,

    Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vi- sion transformer using shifted windows,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 10012–10022, 2021

    2021

  56. [64]

    AION-1: Omnimodal foundation model for astronomi- cal sciences,

    L. Parker, F. Lanusse, J. Shen, O. Liu, T. Hehir, L. Sarra, L. Meyer, M. Bowles, S. Wagner-Carena, H. Qu,et al., “AION-1: Omnimodal foundation model for astronomi- cal sciences,”Advances in Neural Information Processing Systems, vol. 38, pp. 95386–95428, 2026

    2026

  57. [65]

    Gemma: Open mod- els based on Gemini research and technology,

    Gemma Team, T. Mesnard, C. Hardin, R. Dadashi, S. Bhupatiraju, S. Pathak,et al., “Gemma: Open mod- els based on Gemini research and technology,”arXiv preprint arXiv:2403.08295, 2024

    2024 arXiv

  58. [66]

    simulateSearch: High-time resolution data sets simulations for radio telescopes

    G. Hobbs, R. Luo, and L. Toomey, “simulateSearch: High-time resolution data sets simulations for radio telescopes.” Astrophysics Source Code Library, record ascl:2205.025, 2022.https://ascl.net/2205.025

    2022

  59. [67]

    PSRCHIVE and PSRFITS: an open approach to radio pulsar data storage and analysis,

    A. W. Hotan, W. van Straten, and R. Manchester, “PSRCHIVE and PSRFITS: an open approach to radio pulsar data storage and analysis,”Publications of the As- tronomical Society of Australia, vol. 21, no. 3, pp. 302– 309, 2004

    2004

  60. [68]

    Transformers: State-of- the-art natural language processing,

    T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen, C. Ma, Y. Jer- nite, J. Plu, C. Xu, T. Le Scao, S. Gugger, M. Drame, Q. Lhoest, and A. M. Rush, “Transformers: State-of- the-art na...

    2020

  61. [69]

    Verification of forecasts expressed in terms of probability,

    G. W. Brier, “Verification of forecasts expressed in terms of probability,”Monthly Weather Review, vol. 78, no. 1, pp. 1–3, 1950

    1950

  62. [70]

    On calibration of modern neural networks,

    C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” inInternational conference on machine learning, pp. 1321–1330, PMLR, 2017

    2017

  63. [71]

    Note on the sampling error of the dif- ference between correlated proportions or percentages,

    Q. McNemar, “Note on the sampling error of the dif- ference between correlated proportions or percentages,” Psychometrika, vol. 12, no. 2, pp. 153–157, 1947

    1947

  64. [72]

    Bootstrap Methods: Another Look at the Jackknife,

    B. Efron, “Bootstrap Methods: Another Look at the Jackknife,”The Annals of Statistics, vol. 7, no. 1, pp. 1– 26, 1979

    1979

  65. [73]

    Ablation studies in artificial neural networks,

    R. Meyes, M. Lu, C. W. De Puiseau, and T. Meisen, “Ablation studies in artificial neural networks,”arXiv preprint arXiv:1901.08644, 2019

    1901 arXiv

  66. [74]

    Ablation programming for machine learning,

    S. Sheikholeslami, “Ablation programming for machine learning,” Master’s thesis, KTH Royal Institute of Tech- nology, Stockholm, Sweden, 2019.https://urn.kb.se/ resolve?urn=urn:nbn:se:kth:diva-258413

    2019

  67. [75]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. K¨ opf, E. Yang, Z. DeVito, M. Rai- son, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “Pytorch: An imperative style, hig...

    2019

  68. [76]

    Array programming with NumPy,

    C. R. Harris, K. J. Millman, S. J. van der Walt, R. Gom- mers, P. Virtanen, D. Cournapeau, E. Wieser, J. Tay- lor, S. Berg, N. J. Smith, R. Kern, M. Picus, S. Hoyer, M. H. van Kerkwijk, M. Brett, A. Haldane, J. F. del R´ ıo, M. Wiebe, P. Peterson, P. G´ erard-Marchant, K. Shep...

    2020

  69. [77]

    Scikit-learn: Machine learning in Python,

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cour- napeau, M. Brucher, M. Perrot, and ´E. Duchesnay, “Scikit-learn: Machine learning in Python,”Journal of Machine Lear...

    2011

  70. [78]

    Matplotlib: A 2D graphics environment,

    J. D. Hunter, “Matplotlib: A 2D graphics environment,” Computing in Science & Engineering, vol. 9, no. 3, pp. 90–95, 2007. 24

    2007

  71. [79]

    The Astropy project: Sustaining and growing a community-oriented open-source project and the latest major release (v5.0) of the core package,

    Astropy Collaboration, A. M. Price-Whelan, P. L. Lim, N. Earl, N. Starkman, L. Bradley, D. L. Shupe, et al., “The Astropy project: Sustaining and growing a community-oriented open-source project and the latest major release (v5.0) of the core package,”The Astro- physical Journ...

    2022

Pith tools

Reviewed July 9, 2026 · model on record in the stance chip above.