Pith. sign in

REVIEW 5 major objections 6 minor 2 cited by

Finetuning Stellar Spectra Foundation Models with LoRA

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that Low-Rank Adaptation lets a stellar-spectrum foundation model trained on LAMOST and Gaia XP data transfer to DESI with about one hundred labeled stars, cutting [Fe/H] prediction scatter from 0.256 to 0.202 dex.

desk verdict A clear but underpowered empirical study: LoRA fine-tuning beats zero-shot on DESI [Fe/H], but the missing non-LoRA baselines and single-split evaluation leave the method-specific claim unproven. read the letter →

arxiv 2507.20972 v1 pith:2BPOR6WJ submitted 2025-07-28 astro-ph.IM astro-ph.SR

Xiaosheng Zhao , Yuan-Sen Ting , Alexander S. Szalay , Yang Huang This is my paper · ORCID
classification astro-ph.IMastro-ph.SR
keywords low-rankadaptationstellarspectroscopyfoundationmodelsfew-shotlearningSpecCLIPDESIironabundancecross-surveytransfer
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 tries to establish that a stellar-spectrum foundation model can be moved from the surveys it was trained on (LAMOST and Gaia XP) to a different survey (DESI) with very little supervision. The authors fine-tune parts of SpecCLIP with LoRA, a low-rank update trick from large-language-model practice, using only 89 labeled DESI stars for the downstream predictor. The best configuration, LoRA4, reduces the robust scatter $\sigma$ of predicted iron abundance from 0.256 dex (zero-shot) to 0.202 dex on a 396-star test set, while adding only a small fraction of trainable parameters. The result matters because it suggests foundation models for stellar spectra can be adapted to new instruments cheaply, without retraining on large labeled samples or building survey-specific pipelines.

What carries the argument

The central mechanism is Low-Rank Adaptation (LoRA), which leaves a pre-trained network frozen and injects task-specific updates of the form $\Delta W = AB$, where $A$ and $B$ are low-rank matrices, into selected linear layers. The paper applies this to four modules of SpecCLIP: the LAMOST low-resolution-spectrum transformer (LoRA1), the contrastive projection network that maps embeddings into a shared space (LoRA3), and two downstream MLPs, one reading LAMOST embeddings (LoRA2) and one reading Gaia-XP-aligned projected embeddings (LoRA4). A separate design choice carries much of the argument: the foundation and projection modules are fine-tuned on unlabeled DESI spectra, while only the downstream MLPs see the 89 labeled stars, so the reported gains are attributed to adapting representations rather than to supervised fitting of the encoder.

What would settle it

Re-run the LoRA4 and zero-shot MLP2 evaluations on many random recuts of the same 495 APOGEE-matched stars (for example, 50 different 89/9/396 splits) and check whether LoRA4's $\sigma = 0.2023$ beats zero-shot's $\sigma = 0.2560$ in most splits; if the margin collapses or changes sign, the few-shot transfer claim fails. A complementary test is to evaluate the fine-tuned models on DESI stars not used in any split, from a different sky region or epoch, to see whether the gains transfer rather than fitting the test set.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that LoRA fine-tuning makes SpecCLIP, a contrastively pre-trained model built from LAMOST low-resolution spectra and Gaia XP spectra, work on DESI Early Data Release spectra in a few-shot regime. With 89 labeled stars for training, fine-tuning only the downstream MLP (a small neural-network predictor) that reads the Gaia-XP-aligned embeddings (LoRA4) gives $\sigma = 0.2023$ and $R^2 = 0.7937$ for [Fe/H] prediction on the full 396-star test set, versus zero-shot $\sigma = 0.2560$ and $R^2 = 0.7203$ for the corresponding MLP. The paper also finds that adapting the foundation backbone and projection network together (LoRA1 + LoRA3 + LoRA4) beats zero-shot overall, that gains are concentrated in metal-rich stars ([Fe/H] $\geq -1$), and that all fine-tuned variants perform worse than zero-shot on the small metal-poor subset. The authors interpret this as evidence that Gaia XP knowledge embedded in the pre-trained model helps adaptation to DESI, and that LoRA is a lightweight route to survey transfer.

Load-bearing premise

The load-bearing premise is that one random split of the 495 labeled stars into 89 training, 9 validation, and 396 test stars gives a reliable ranking of the configurations; if sampling noise is large, the reported margin of LoRA4 over zero-shot could disappear.

Editorial extensions

If this is right

  • A foundation model trained on one set of surveys can be redeployed to a new survey with around 100 labeled stars and a negligible parameter budget (about 0.3% of the backbone).
  • For DESI at least, the best transfer path is to fine-tune the predictor on Gaia-XP-aligned embeddings rather than the LAMOST-native embeddings, which points to cross-modal alignment as the survey-invariant representation.
  • Jointly adapting the foundation and projection layers improves over zero-shot on the full test set and over MLP-only tuning on metal-rich stars, but the gains are not monotonic when the whole architecture is tuned.
  • Metal-poor stars remain a failure mode: every LoRA-tuned variant degrades relative to zero-shot, so the current few-shot recipe is not yet reliable in sparse label regimes.
  • Training cost is small enough (seconds to minutes per experiment on one GPU) that per-survey fine-tuning is practical at survey scale.

Reading between the lines

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

  • If this transfer recipe holds for other instruments, spectroscopic surveys could homogenize stellar parameters by fine-tuning one foundation model per instrument on the small set of stars observed by both surveys, rather than retraining from scratch.
  • The LoRA4 result suggests a testable hypothesis: the Gaia-XP-aligned embedding space is more invariant across instruments than the LAMOST-native embedding space; one could probe this by repeating the experiment on, say, GALAH or SDSS-V spectra and checking whether the projected MLP again wins.
  • Because the reported ranking rests on one random split, a natural stress test is to repeat the 89/9/396 split many times and report distributions of $\sigma$ and $R^2$; the paper does not do this, and its point estimates could shift.
  • A direct extension would be to oversample metal-poor stars in the 89-star training set (or add label-balanced sampling) and see whether the metal-poor degradation vanishes; if it does not, the bottleneck is representation rather than label scarcity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper applies Low-Rank Adaptation (LoRA) to SpecCLIP, a spectral foundation model pre-trained on LAMOST and Gaia XP spectra, in order to adapt it to DESI Early Data Release stellar spectra for iron abundance ([Fe/H]) estimation. Using 89 APOGEE-cross-matched DESI spectra for training, 9 for validation, and 396 for testing, the authors compare zero-shot SpecCLIP predictors with several LoRA configurations that adapt different modules: the LAMOST backbone (LoRA1), the projection network (LoRA3), and the two downstream MLPs (LoRA2 and LoRA4). The best configuration, LoRA4, which adapts only the MLP on the projected Gaia-XP-aligned embeddings, achieves sigma=0.2023 and R2=0.7937 on the full test set, compared with sigma=0.2560 and R2=0.7203 for the corresponding zero-shot predictor. The paper reports strong gains on the metal-rich subset and degraded performance on the metal-poor subset, and concludes that LoRA is a lightweight and effective strategy for cross-survey transfer.

Significance. If the central claim is established, the result would be practically valuable: it would show that a contrastively pre-trained spectral foundation model can be transferred to a new instrument with roughly one hundred labeled spectra and a small trainable-parameter budget. The paper uses real DESI EDR and APOGEE DR17 data, labels from an independent catalog, and a transparent comparison of several adaptation configurations, and the reported runtimes on a single V100 are useful practical information. The main limitation is that the paper is purely empirical and, as written, does not isolate the effect of the low-rank constraint from the effect of supervised adaptation: every comparison is against a frozen zero-shot baseline, and all metrics are single point estimates from one train/validation/test split. Consequently, the contribution is currently more of a promising demonstration than a fully supported methodological claim.

major comments (5)
  1. [§1, §4, Table 1] The central claim that 'LoRA enables effective few-shot learning on DESI data' is not supported by the presented comparisons. Table 1 compares LoRA variants only against frozen zero-shot predictors, but LoRA2 and LoRA4 modify the downstream MLP weights, so the improvement from sigma=0.2560 (zero-shot MLP2) to sigma=0.2023 (LoRA4) could, in principle, be obtained by any supervised retraining of the MLP on the 89 DESI labels, with the low-rank constraint being incidental. I request control experiments: (i) fully fine-tuning the same MLP without the low-rank constraint on the same labels; (ii) training a new MLP or a linear probe on the frozen LRS and projected SpecCLIP embeddings; and (iii) if the claim concerns the backbone, fully fine-tuning the projection network or transformer without LoRA. Without such controls, the title and abstract overstate what the experiments demonstrate.
  2. [§3, Table 1] All reported metrics are point estimates from a single random 89/9/396 split. With only 89 labeled training spectra and 396 test spectra, the differences that drive the paper's conclusions, for example LoRA4 sigma=0.2023 versus LoRA1+LoRA3+LoRA4 sigma=0.2297, or LoRA4 R2=0.7937 versus LoRA1+LoRA2 R2=0.7719, could easily change under sampling noise. The authors should report at least bootstrap confidence intervals on sigma and R2, and preferably repeated random splits with different seeds or a cross-validation scheme; the 9-star validation set also makes any hyperparameter selection fragile.
  3. [§4, §5] There is an internal inconsistency between the text and the table. Section 4 states that 'further gains can be achieved by jointly adapting the foundation model,' and Section 5 repeats that 'further gains can be achieved by jointly tuning the foundation model layers,' but Table 1 shows that LoRA1+LoRA3+LoRA4 is worse than LoRA4 alone on the full test set (sigma=0.2297 versus 0.2023; R2=0.7801 versus 0.7937) and on the metal-rich subset (sigma=0.1851 versus 0.1621). This claim should be corrected, or the paper should explain why the additional adaptation degrades performance.
  4. [§2.2] The fine-tuning protocol is under-specified, which prevents replication. The paper does not state the objective used for LoRA1 and LoRA3 on the unlabeled DESI spectra (masked modeling, contrastive loss, or a combination), nor the optimizer, learning rate, batch size, number of epochs, early-stopping criterion, or how the reported ranks and scaling factors were chosen. Because the paper compares multiple fine-tuning configurations, the reader cannot tell whether all configurations were trained under comparable and appropriate conditions.
  5. [§1, §4] The abstract's claim that performance 'benefits from Gaia XP knowledge embedded in the pre-trained model' is not isolated by the experiments. LoRA4 differs from LoRA2 in both the input embedding (projected versus LRS embeddings) and the initial MLP, while zero-shot MLP2 also differs from zero-shot MLP1 in the same way. A matched comparison that trains the same MLP architecture on both embedding types with identical hyperparameters, or an ablation that removes the XP-aligned projection, is needed before attributing the gain to Gaia XP knowledge.
minor comments (6)
  1. [§1] The phrase 'learn/arange both shared and non-shared information' appears to contain a typo; it should likely read 'learn/arrange' or 'learn and arrange'.
  2. [§2.1] The sentence 'This unsupervised pre-training retrain the most important information' should read 'retains the most important information'.
  3. [§3] The footnote explaining that 86 training and 78 validation spectra were ultimately used instead of the intended 100 each is useful, but the final counts should be stated in the body of Section 3 rather than only in a footnote.
  4. [Figure 1] The caption phrase 'and optionally the LRS foundation model' is ambiguous about which configuration includes the optional component; labeling each branch with the exact configuration names from Section 2.2 would improve clarity.
  5. [Table 1] Since parameter efficiency is a stated motivation, the table would be more informative with a column reporting the number of trainable parameters for each configuration.
  6. [§3] The paper treats APOGEE DR17 [Fe/H] as ground truth but does not discuss label uncertainties; because the reported test-set sigmas (0.16-0.25 dex) are not far above typical APOGEE abundance uncertainties, a brief discussion of this contribution to the scatter would be helpful.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the transfer claim is tested against external APOGEE labels; SpecCLIP self-citation is the object under study, not a circularity.

full rationale

No circular step meets the evidentiary bar. The central comparison (Section 4, Table 1) uses APOGEE DR17 [Fe/H] labels that are external to SpecCLIP and to the LoRA procedure; LoRA4's sigma=0.2023/R2=0.7937 versus zero-shot MLP2's sigma=0.2560/R2=0.7203 is an empirical, falsifiable result, not a quantity restored by construction. Fine-tuning uses 89 DESI labels for LoRA2/LoRA4 and unlabeled DESI spectra for LoRA1/LoRA3, and no fitted parameter is later reported as a prediction. The cited SpecCLIP paper (Zhao et al., 2025, arXiv:2507.01939) supplies the model being adapted; citing the model one is fine-tuning is structural, and the prior work has its own external evaluations rather than being invoked as an unverified uniqueness theorem. Legitimate concerns in this manuscript are not circular: Section 3's single random split (89/9/396) with point-estimate metrics in Table 1 is a statistical robustness issue; the absence of a fully fine-tuned or linear-probe control leaves the title claim 'LoRA enables' not fully isolated from ordinary supervised head adaptation; and Section 4's 'further gains can be achieved by jointly adapting the foundation model' contradicts Table 1 (LoRA1+LoRA3+LoRA4 sigma=0.2297 vs LoRA4 alone sigma=0.2023). These are experimental-design or consistency criticisms, not equivalences between inputs and outputs.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a chain of empirical modeling assumptions: the DESI-to-LAMOST interpolation preserves abundance information, APOGEE labels are accurate ground truth, and the frozen SpecCLIP embeddings transfer across instruments. The hand-chosen LoRA hyperparameters and sample split sizes are additional free choices that are not swept or justified. No new physical entities are introduced.

free parameters (5)
  • LoRA rank r per module = r=4 (LoRA1), r=16 (LoRA3), r=8 (LoRA2/LoRA4)
    Chosen by hand in Section 2.2; no rank sweep is reported, so the configuration ranking could depend on these choices.
  • LoRA scaling factor alpha per module = alpha=8 (LoRA1), alpha=32 (LoRA3), alpha=16 (LoRA2/LoRA4)
    Chosen by hand in Section 2.2; no sensitivity analysis is given.
  • Number of unlabeled DESI spectra for LoRA1/LoRA3 = 164 (86 train, 78 validation)
    Selected after excluding the 495 labeled stars; footnote 2 says the intended 100+100 was reduced by preprocessing constraints, so the amount is somewhat arbitrary.
  • Labeled split sizes = 89 train, 9 validation, 396 test
    A single random split with no seed reported (Section 3); all downstream metrics depend on this one split.
  • Quality filter thresholds = SNR > 20 (main sample), SNR > 50 (unlabeled subset), STAR, RVS_WARN=0, PRIMARY
    Adopted in Section 3 without sensitivity checks; could change which stars enter the test set.
assumptions (4)
  • domain assumption Interpolating DESI spectra onto the LAMOST 400-560 nm wavelength grid preserves the information needed for [Fe/H] prediction.
    Section 3 asserts the standardization 'ensures compatibility', but the paper does not test whether the truncated wavelength range retains the abundance information.
  • domain assumption APOGEE DR17 [Fe/H] labels are accurate enough to serve as ground truth for the DESI test set.
    Section 3 cross-matches with APOGEE; if label errors are comparable to the reported scatter differences, the ranking of configurations may change.
  • domain assumption Frozen SpecCLIP embeddings trained on LAMOST and Gaia XP generalize to DESI spectra after interpolation.
    This is the core transfer premise of Sections 2.1 and 2.2; no experiment verifies it except the zero-shot baseline itself.
  • domain assumption LoRA low-rank updates are sufficient to capture the domain shift from LAMOST to DESI.
    The paper uses LoRA without comparing to full fine-tuning (Section 2.2), so the sufficiency of the low-rank constraint is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Finetuning Stellar Spectra Foundation Models with LoRA." pith.science (2026). https://pith.science/paper/2BPOR6WJ

@misc{pith2026250720972,
  author       = {Pith},
  title        = {Pith review of: Finetuning Stellar Spectra Foundation Models with LoRA},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2BPOR6WJ}},
  note         = {Machine review of arXiv:2507.20972}
}
read the original abstract

Foundation models are beginning to impact stellar spectroscopy, where spectra encode rich physical information in a structured, language-like form. A key challenge is adapting these models across heterogeneous surveys with differing resolution and coverage. We apply Low-Rank Adaptation (LoRA) to fine-tune SpecCLIP--a contrastively pre-trained model on LAMOST and Gaia XP spectra--for downstream tasks on DESI Early Data Release (EDR) spectra. We show that LoRA enables few-shot learning on DESI, with performance varying by fine-tuned module and benefiting from Gaia XP knowledge embedded in the pre-trained model. Our results demonstrate that LoRA provides a lightweight and effective strategy for extending spectral foundation models to new instruments and survey domains.

Figures

Figures reproduced from arXiv: 2507.20972 by the authors.

Figure 1
Figure 1. Fine-tuning the SpecCLIP model with LoRA. Gray rounded blocks represent pre-trained modules: the LAMOST LRS and Gaia XP foundation models (trained separately on their respective modalities), and the projection networks trained with contrastive loss on paired projected embeddings. Two downstream MLPs, previously trained on LRS-based iron abundance labels, are also included. In the fine-tuning stage, a small number of… view at source ↗
Figure 2
Figure 2. Scatter plots of predicted versus APOGEE iron abun￾dance ([Fe/H]) for all fine-tuning configurations, with point den￾sity indicated by the viridis colormap. The dashed black line indicates the ideal 1:1 relation, and the gray dotted lines mark ±0.2 dex deviations. An inset zooms into the ±0.3 dex region for closer inspection. LoRA fine-tuning improves overall predictive accuracy compared to zero-shot baselines, part… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Generalization from Low- to Moderate-Resolution Spectra with Neural Networks for Stellar Parameter Estimation: A Case Study with DESI

    astro-ph.SR 2026-02 conditional novelty 6.0 of 10

    Pre-trained MLPs on LAMOST low-resolution spectra generalize to DESI medium-resolution spectra for [Fe/H] and [α/Fe], outperforming the DESI SP pipeline in zero-shot and improving with modest fine-tuning.

  2. Foundation Models for Astrophysics

    astro-ph.IM 2026-08 conditional novelty 3.0 of 10

    Astronomical 'foundation models' largely reuse transformers and self-supervised pretraining, but evidence of transfer to new instruments, populations, or tasks remains rare; the paper argues such evidence, not archite...

Reference graph

Works this paper leans on

19 extracted references · 6 linked inside Pith · cited by 2 Pith papers

  1. [7]

    2021.3057097

    doi: 10.1109/MCSE. 2021.3057097. Juneau, S., Jacques, A., Pothier, S., Bolton, A. S., Weaver, B. A., Pucha, R., McManus, S., Nikutta, R., and Olsen, K. SPARCL: SPectra Analysis and Retrievable Catalog Lab. arXiv e-prints, art. arXiv:2401.05576, January

    arXiv 2021

  2. [10]

    Ness, M., Hogg, D

    doi: 10.1093/mnras/stae1842. Ness, M., Hogg, D. W., Rix, H. W., Ho, A. Y . Q., and Zasowski, G. The Cannon: A data-driven approach to Stellar Label Determination. ApJ, 808(1):16, July

    doi:10.1093/mnras/stae1842

  3. [13]

    Radford, A., Kim, J

    doi: 10.1093/ mnras/stae1450. Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. Learning Transfer- able Visual Models From Natural Language Supervision. arXiv e-prints , art. arXiv:2103.00020, February

    arXiv

  4. [14]

    R´o˙za´nski, T., Ting, Y .-S., and Jabło´nska, M

    doi: 10.48550/arXiv.2103.00020. R´o˙za´nski, T., Ting, Y .-S., and Jabło´nska, M. Transformer- Payne: Enhancing Spectral Emulation Accuracy and Data Efficiency by Capturing Long-range Correlations. ApJ, 980(1):66, February

    arXiv doi:10.48550/arxiv.2103.00020

  5. [17]

    Xiang, M

    doi: 10.3847/1538-4365/ab5364. Xiang, M. S., Liu, X. W., Yuan, H. B., Huang, Y ., Huo, Z. Y ., Zhang, H. W., Chen, B. Q., Zhang, H. H., Sun, N. C., Wang, C., Zhao, Y . H., Shi, J. R., Luo, A. L., Li, G. P., Wu, Y ., Bai, Z. R., Zhang, Y ., Hou, Y . H., Yuan, H. L., Li, G. W., and Wei, Z. The LAMOST stellar parameter pipeline at Peking University - LSP3. M...

    doi:10.3847/1538-4365/ab5364

  6. [18]

    Zhang, M., Xiang, M., Ting, Y .-S., Wang, J., Li, H., Zou, H., Nie, J., Mou, L., Wu, T., Wu, Y ., and Liu, J

    doi: 10.1093/mnras/stu2692. Zhang, M., Xiang, M., Ting, Y .-S., Wang, J., Li, H., Zou, H., Nie, J., Mou, L., Wu, T., Wu, Y ., and Liu, J. Determining Stellar Elemental Abundances from DESI Spectra with the Data-driven Payne. ApJS, 273(2):19, August

    doi:10.1093/mnras/stu2692

  7. [19]

    Zhao, X., Huang, Y ., Xue, G., Kong, X., Liu, J., Tang, X., Beers, T

    doi: 10.3847/1538-4365/ad51dd. Zhao, X., Huang, Y ., Xue, G., Kong, X., Liu, J., Tang, X., Beers, T. C., Ting, Y .-S., and Luo, A.-L. SpecCLIP: Aligning and Translating Spectroscopic Measurements for Stars. arXiv e-prints , art. arXiv:2507.01939, July

    arXiv

  8. [20]

    doi: 10.48550/arXiv.2507.01939. 7

    doi:10.48550/arxiv.2507.01939

Show all 19 references
  1. [1983]

    J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., and Chen, W

    Hu, E. J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., and Chen, W. LoRA: Low-Rank Adap- tation of Large Language Models. arXiv e-prints , art. arXiv:2106.09685, June

    arXiv

  2. [2009]

    Koposov, S

    doi: 10.1051/0004-6361/200811467. Koposov, S. E., Allende Prieto, C., Cooper, A. P., Li, T. S., Beraldo e Silva, L., Kim, B., Carrillo, A., Dey, A., Manser, C. J., Nikakhtar, F., Riley, A. H., Rock- osi, C., Valluri, M., Aguilar, J., Ahlen, S., Bailey, S., Blum, R., Brooks, D....

    doi

  3. [2014]

    URL https://doi.org/ 10.1117/12.2057445

    doi: 10.1117/12.2057445. URL https://doi.org/ 10.1117/12.2057445. Gaia Collaboration, Montegriffo, P., Bellazzini, M., De Angeli, F., Andrae, R., Barstow, M. A., Bossini, D., Bra- gaglia, A., Burgess, P. W., Cacciari, C., Carrasco, J. M., Chornay, N., Delchambre, L., Evans, D....

    doi

  4. [2015]

    Nikutta, R., Fitzpatrick, M., Scott, A., and Weaver, B

    doi: 10.1088/0004-637X/808/1/16. Nikutta, R., Fitzpatrick, M., Scott, A., and Weaver, B. Data lab—a community science platform. Astronomy 6 LoRA on Stellar Spectra Foundation Models and Computing , 33:100411,

    doi

  5. [2018]

    1810.04805

    doi: 10.48550/arXiv. 1810.04805. Fitzpatrick, M. J., Olsen, K., Economou, F., Stobie, E. B., Beers, T. C., Dickinson, M., Norris, P., Saha, A., Sea- man, R., Silva, D. R., Swaters, R. A., Thomas, B., and 5 LoRA on Stellar Spectra Foundation Models Valdes, F. The NOAO Data Labo...

    doi arXiv

  6. [2019]

    Xiang, M., Ting, Y .-S., Rix, H.-W., Sandford, N., Buder, S., Lind, K., Liu, X.-W., Shi, J.-R., and Zhang, H.-W

    doi: 10.3847/1538-4357/ ab2331. Xiang, M., Ting, Y .-S., Rix, H.-W., Sandford, N., Buder, S., Lind, K., Liu, X.-W., Shi, J.-R., and Zhang, H.-W. Abundance Estimates for 16 Elements in 6 Million Stars from LAMOST DR5 Low-Resolution Spectra. ApJS, 245 (2):34, December

    doi

  7. [2020]

    doi: https://doi.org/10.1016/j.ascom.2020.100411

    ISSN 2213-1337. doi: https://doi.org/10.1016/j.ascom.2020.100411. URL https://www.sciencedirect.com/ science/article/pii/S2213133720300652. Parker, L., Lanusse, F., Golkar, S., Sarra, L., Cranmer, M., Bietti, A., Eickenberg, M., Krawezik, G., McCabe, M., Morel, R., Ohana, R., ...

    2020

  8. [2021]

    doi: 10.48550/arXiv.2106. 09685. Juneau, S., Olsen, K., Nikutta, R., Jacques, A., and Bailey, S. Jupyter-enabled astrophysical analysis using data- proximate computing platforms. Computing in Science & Engineering, 23(2):15–25,

    doi

  9. [2022]

    Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K

    doi: 10.3847/1538-4365/ac4414. Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. BERT: Pre-training of Deep Bidirectional Transform- ers for Language Understanding. arXiv e-prints , art. arXiv:1810.04805, October

    arXiv

  10. [2023]

    Hoaglin, D

    doi: 10.1051/0004-6361/202243709. Hoaglin, D. C., Mosteller, F., and Tukey, J. W.Understand- ing Robust and Exploratory Data Analysis . John Wiley & Sons, New York,

    doi

  11. [2024]

    Koleva, M., Prugniel, P., Bouchard, A., and Wu, Y

    doi: 10.48550/arXiv.2401.05576. Koleva, M., Prugniel, P., Bouchard, A., and Wu, Y . ULySS: a full spectrum fitting package. A&A, 501(3):1269–1279, July

    doi arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.