Pith. sign in

REVIEW 3 major objections 6 minor 28 references

Transfer Learning and Mixup for Fine-Grained Few-Shot Fungi Classification

T0 review · 3 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 This paper claims that freezing a plant-pretrained vision transformer, rebalancing the few-shot training set with weighted sampling, and applying feature-level Mixup beats the official BioCLIP baselines on FungiCLEF 2025, with tuned Mixup…

desk verdict A transparent competition working note with a useful practical recipe, but the marquee 4.27% mixup gain is post-hoc: the alpha was effectively selected after seeing private scores, so that specific number should not be taken at face value. read the letter →

arxiv 2507.08248 v1 pith:MVO37AGQ submitted 2025-07-11 cs.CV cs.IRcs.LG

Jason Kahei Tam , Murilo Gustineli , Anthony Miyaguchi This is my paper
classification cs.CVcs.IRcs.LG
keywords few-shotlearningfine-grainedvisualcategorizationfungiclassificationtransfermixupweightedsamplingvisiontransformersLifeCLEF
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

The paper argues that in the extreme few-shot regime of FungiCLEF 2025, the effective recipe is cheap: freeze a vision transformer pretrained on a neighbouring domain, rebalance the small training set with inverse-frequency sampling, and interpolate embedding vectors between classes with Mixup before a linear classifier. The authors report that this pipeline, with Mixup strength $\alpha$ tuned to 1.20 or 1.45, reaches 47.3% top-5 accuracy on the private test set and beats both official BioCLIP baselines, which score 26.6% and 24.7%. They attribute the largest single gain, 4.27 percentage points, to Mixup with a tuned $\alpha$ operating on features rather than raw images. They also find that adding text metadata, multi-objective GradNorm training, or generative vision-language models decreases accuracy, reinforcing the conclusion that the gains come from domain-specific embeddings and simple balancing.

What carries the argument

The carrying mechanism is Mixup applied to the frozen image embeddings rather than to raw pixels: each training batch is blended with a shuffled copy, producing $\tilde{x} = \lambda x_i + (1-\lambda) x_j$ with $\lambda \sim \mathrm{Beta}(\alpha,\alpha)$, and the cross-entropy loss is mixed accordingly, $\mathcal{L}_{\mathrm{Mixup}} = \lambda \mathcal{L}(f(\tilde{x}), y_i) + (1-\lambda)\mathcal{L}(f(\tilde{x}), y_j)$. Because the embeddings come from a PlantCLEF 2024 model already pretrained on a large botanical image collection, the classifier only needs to separate classes near each other in that feature space. Weighted sampling (with weights set to inverse class frequency) and the tuned $\alpha$ values work together to keep rare classes from being ignored; the paper's ablation isolates Mixup as the larger contributor.

What would settle it

Rerun the PlantCLEF-embedding linear classifier on a fixed stratified split of the FungiTastic few-shot data with several Mixup alphas (for example 0.1, 0.4, 1.2, 1.45, 2.0) and multiple random seeds; if the best alpha changes across splits and seeds, the claim that tuned Mixup is the single largest contributor to the private gain is not supported.

Watch

Extended reading notes

Core claim

The central claim is that feature-level Mixup, applied to precomputed PlantCLEF 2024 embeddings and paired with inverse-frequency weighted sampling, is the deciding component in a few-shot fungi classifier. On the competition's private test set, the post-competition configuration with $\alpha=1.45$ achieves 47.347% top-5 accuracy and the competition submission (Mixup $\alpha=2.0$ with weighted sampling) achieves 45.407%; both exceed the BioCLIP+FAISS prototype baseline (26.649%) and the BioCLIP+FAISS nearest-neighbour baseline (24.708%). The paper states that tuned Mixup alone accounts for a 4.27% private-score improvement, that weighted sampling gives a smaller gain, and that metadata-plus-caption text embeddings and GradNorm multi-objective training hurt performance.

Load-bearing premise

The load-bearing premise is that the ranking of Mixup alpha values on the public leaderboard transfers to the private test set; the paper's own results show this transfer is imperfect, so the reported 4.27% Mixup gain depends on that selection step.

Editorial extensions

If this is right

  • A domain-neighbour pretrained embedding (plant images) transfers at least as well as a generic self-supervised ViT and a fungi-pretrained ViT for few-shot fungus identification.
  • Mixup alpha values around 1.2-1.45, chosen by public leaderboard probing, outperform the values recommended for input-level Mixup, suggesting feature-level interpolation behaves differently.
  • Adding text captions and metadata, multi-objective GradNorm, or generative LLM prompting reduces accuracy, so simple vision-only pipelines are hard to beat at this data scale.
  • The final system's private score (45.4-47.3%) and rank (35/74) show the gap to top systems is large, implying more work on architecture and metadata is needed.

Reading between the lines

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

  • The alpha values were selected on the public leaderboard, and the paper's own results show the ranking shifts on the private set: alpha 1.20 with weighted sampling has the best public score (50.884) but a lower private score (46.830) than alpha 1.45 alone (47.347), so the headline 4.27% gain is tied to a selection procedure the paper does not validate on private data.
  • Feature-level Mixup on frozen embeddings is conceptually close to Manifold Mixup; applying Mixup at a randomly chosen layer of a small trainable network might smooth class boundaries further and reduce the volatility the paper sees in rare-class accuracy.
  • The failed metadata experiments may be reversible: the paper used all available fields including district, countryCode, and hasCoordinate, and itself suggests that restricting to substrate, location, and date could recover the positive metadata effect reported by the FungiTastic benchmark.
  • Because the final classifier is a linear layer on 768-dimensional precomputed embeddings, the whole pipeline is cheap to rerun and ensemble; a testable extension is to average predictions across alpha values or across the PlantCLEF and FungiTastic ViT embeddings.
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

3 major / 6 minor

Summary. The paper reports the DS@GT entry to FungiCLEF 2025, a few-shot fine-grained fungi classification challenge. The authors use a frozen ViT backbone (the PlantCLEF 2024 DINOv2-based model) to extract embeddings, train a linear classifier with weighted random sampling and feature-level mixup, and ablate several variants including text embeddings, GradNorm multi-objective training, and zero-shot prompting of commercial vision-language models. On the competition's private test set, their best in-competition submission (mixup alpha=2 with weighted sampling) achieves 45.407% top-5 accuracy, beating the BioCLIP baselines (26.649% and 24.708%); post-competition tuning reaches 47.347% with mixup alpha=1.45. The paper concludes that domain-specific pretraining and balanced sampling help and that mixup with a tuned alpha is the single most effective technique.

Significance. If the central attribution claims were supported, the paper would provide useful evidence that feature-level mixup and class-balanced sampling transfer to few-shot fine-grained fungi recognition, and that generic zero-shot VLM identification is not yet competitive. Strengths include the evaluation against official held-out private test labels, comparison with external BioCLIP baselines, an unusually complete ablation table, and public code. The manuscript is honest about the modest leaderboard rank (35/74) and about the volatility of rare-class accuracy. However, the headline attribution of the 4.27% gain to a tuned alpha is weakened by the fact that the winning alpha was selected after observing private scores; the public leaderboard cannot distinguish alpha=1.20 from alpha=1.45, and all comparisons come from single runs without error bars.

major comments (3)
  1. [Section 4.2, Table 5, Figure 5] The central claim that 'Mixup with a tuned α is the single technique with the greatest positive impact with an increase of 4.27% on the private score' is not a pre-specified comparison. Section 3.1.3 states that α=1.20 and α=1.45 tied for the highest public score (52.654), and Table 5 shows their private scores differ by 2.587 points (47.347 vs 44.760). Since the public leaderboard cannot distinguish these two configurations, the choice of α=1.45 as the 'tuned' value is justified only after seeing the private labels. The instability is further demonstrated by the weighted-sampling rows: α=1.45 + weighted sampling has the best public score in the table (53.982) but a lower private score (46.054) than α=1.20 + weighted sampling (46.830). With a single run on a 1,911-image test set, the reported 4.27% improvement is therefore an observed maximum over private comparisons, not a validated effect; the paper should either select α using the held-out validation split or repeated seeds, or reframe the claim as an exploratory post-competition observation.
  2. [Table 5 and Section 3.1.3] The ablation row 'w/ Mixup (α=2.00)* (Competition)' was trained with 10 epochs while all other rows, including the baseline, used 50 max epochs. The drop from the baseline in that row (46.460 vs 48.672 public; 40.750 vs 43.078 private) is therefore not attributable to Mixup alone; it may simply reflect insufficient training. The post-competition rows with 50 epochs are the clean comparison, and the paper should either retrain the α=2.00 configuration at 50 epochs or clearly mark this row as not comparable.
  3. [Table 5 and Section 4.2] All results are single runs with no error bars or significance tests. Since the test set has 1,911 images and several configuration differences are smaller than 1.5 points (e.g., baseline vs weighted sampling: 43.078 vs 44.372; α=1.20+weighted vs α=1.45+weighted: 46.830 vs 46.054), the paper's ordering of techniques may be within run-to-run noise. Given the reported early-stopping procedure, retraining with different seeds is inexpensive; reporting mean ± standard deviation over at least three seeds would materially strengthen the ablation.
minor comments (6)
  1. [Table 6] In the row 'Public PlantCLEF + Mixup ( α = 2.00) + Weighted Sampling (Competition)', the opening parenthesis before 'α' is missing; using the same naming convention as Table 5 would also improve readability.
  2. [Section 3.2.3] The prompt block is referenced as 'listing 3.2.3' but has no caption; please add a listing caption or change the cross-reference.
  3. [Section 3.2.1] The phrase 'categories present in the test metadata file' is confusing because Section 1.1 states the test set has no taxonomic ranks; clarify whether this refers to class names from the training metadata or to some other file, since using test labels would be a leakage concern.
  4. [Table 4] PlantCLEF 2024 and FungiTastic ViT tie at 48.672 public accuracy, yet the text says PlantCLEF was selected; please state the tie-breaking criterion (e.g., private score or training cost).
  5. [Figure 5] The y-axis label 'Top-K Accuracy' does not state K; since the evaluation metric is top-5 accuracy, the figure should label the axis 'Top-5 Accuracy'.
  6. [References] Reference [11] cites a Medium blog post for WeightedRandomSampler; citing the official PyTorch documentation or the original source would be more appropriate for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central comparisons are evaluated on an independent private test set.

full rationale

The paper's derivation chain is empirical and externally anchored. The central quantitative claim—that Mixup with a tuned alpha improves private top-5 accuracy by 4.27%—is read directly from Table 5 (47.347 vs. 43.078); no equation defines one quantity in terms of another, and the private test set is independent of the public leaderboard used for alpha selection. The self-citation [7] is used only as motivation for trying DINOv2 and multi-objective losses; DINOv2's choice is independently benchmarked in Table 4, and the multi-objective experiments are reported as failures, so the citation is not load-bearing. The alpha-selection procedure (public leaderboard) could inflate the reported gain through selection bias, but that is a statistical robustness concern, not circularity: the public scores do not by construction determine the private scores, and the paper does not rename a fitted parameter as a prediction. No circular step is present.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on one data-tuned hyperparameter (mixup alpha) and on the implicit assumptions that the competition labels are reliable, that pre-trained embeddings carry enough fine-grained signal, and that public-leaderboard selection transfers to the private test set. No new entities are introduced; the generative AI part uses existing API models.

free parameters (2)
  • Mixup alpha (alpha) = 1.20 and 1.45 (also 2.00 in competition)
    Selected by highest public leaderboard score in Section 4.2 and Figure 5; directly determines the reported 4.27% private improvement.
  • Top-k candidate cutoff for LLM prompting = 20
    Chosen ad hoc for the generative AI approach; not central to the vision result.
assumptions (4)
  • domain assumption The FungiTastic dataset labels (category_id) are ground truth with negligible noise.
    All accuracy numbers and conclusions rest on label correctness; the paper does not audit labels beyond noting one corrupted image.
  • domain assumption Pre-trained embeddings (PlantCLEF 2024, DINOv2, FungiTastic) encode enough discriminative visual information for a linear classifier to succeed on this fine-grained task.
    The method relies on transfer learning without fine-tuning the backbone; if embeddings lack fine-grained detail, the approach fails regardless of classifier.
  • ad hoc to paper The public leaderboard subset is a representative sample of the private test distribution for hyperparameter selection.
    Alpha values are chosen by public scores in Section 4.2; private results show imperfect rank transfer, making this assumption load-bearing for the reported effect size.
  • domain assumption Top-5 accuracy on the competition test set is an appropriate measure of generalization for fungi identification.
    The paper uses the competition metric throughout; this is given by the task rather than justified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Transfer Learning and Mixup for Fine-Grained Few-Shot Fungi Classification." pith.science (2026). https://pith.science/paper/MVO37AGQ

@misc{pith2026250708248,
  author       = {Pith},
  title        = {Pith review of: Transfer Learning and Mixup for Fine-Grained Few-Shot Fungi Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MVO37AGQ}},
  note         = {Machine review of arXiv:2507.08248}
}
read the original abstract

Accurate identification of fungi species presents a unique challenge in computer vision due to fine-grained inter-species variation and high intra-species variation. This paper presents our approach for the FungiCLEF 2025 competition, which focuses on few-shot fine-grained visual categorization (FGVC) using the FungiTastic Few-Shot dataset. Our team (DS@GT) experimented with multiple vision transformer models, data augmentation, weighted sampling, and incorporating textual information. We also explored generative AI models for zero-shot classification using structured prompting but found them to significantly underperform relative to vision-based models. Our final model outperformed both competition baselines and highlighted the effectiveness of domain specific pretraining and balanced sampling strategies. Our approach ranked 35/74 on the private test set in post-completion evaluation, this suggests additional work can be done on metadata selection and domain-adapted multi-modal learning. Our code is available at https://github.com/dsgt-arc/fungiclef-2025.

Figures

Figures reproduced from arXiv: 2507.08248 by the authors.

Figure 1
Figure 1. Example of intra-species variation - Species: Comatricha alta CLEF 2025: Working Notes, 9 – 12 September 2025, Madrid, Spain *Corresponding author. $ jtam30@gatech.edu (J. K. Tam); murilogustineli@gatech.edu (M. Gustineli); acmiyaguchi@gatech.edu (A. Miyaguchi)  0009-0005-0853-0414 (J. K. Tam); 0009-0003-9818-496X (M. Gustineli); 0000-0002-9165-8718 (A. Miyaguchi) © 2025 Copyright for this paper by its authors. Use… view at source ↗
Figure 2
Figure 2. Example of inter-species similarity and variation from three distinct families. Species left to right: Psathyrella citerinii, Inocybe assimilata, and Entoloma favrei [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Distribution of classes in training and validation datasets 1.1. Dataset Overview The dataset provided for the competition is the few-shot subset of the FungiTastic dataset, a collection of fungal records continuously collected over a twenty-year span [5]. Each observation in the dataset contains associated images, metadata, and vision language model (VLM), Molmo [6], generated caption. The metadata contains informa… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Pipeline of the benchmark methodology 3.1.1. Dataset Preparation We pre-computed both the image and text embeddings and stored them in parquet files for a modular experimentation workflow. We encountered "Premature End of JPEG file" error when reading images because so…
Figure 5
Figure 5. Figure 5: Effect of Mixup Alpha on Top-K Accuracy 3.2. Additional Methodologies 3.2.1. Text Embeddings We used ModernBERT-Large[21], a state-of-the-art BERT variant optimized for efficiency, to compute 1024-dimensional text embeddings. We concatenated text from categories presen…
Figure 6
Figure 6. Figure 6: Class Frequency vs Top-5 Accuracy on Validation Set 5.2. Image Embeddings Generation Among the different models evaluated for image embedding generation, the PlantCLEF 2024 and FungiTastic ViT models performed the best ( [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 10 canonical work pages

  1. [1]

    Lücking, M

    R. Lücking, M. Aime, B. Robbertse, et al., Unambiguous identification of fungi: where do we stand and how accurate and precise is fungal dna barcoding?, IMA Fungus (2020)

    work page 2020

  2. [2]

    URL: https://kaggle.com/ competitions/fungi-clef-2025

    Klarka, picekl, FungiCLEF 2025 @ CVPR-FGVC & LifeCLEF, 2025. URL: https://kaggle.com/ competitions/fungi-clef-2025

    work page 2025

  3. [3]

    A. Joly, L. Picek, S. Kahl, H. Goëau, L. Adam, C. Botella, M. Servajean, D. Marcos, C. Leblanc, T. Larcher, J. Matas, K. Janoušková, V. Čermák, K. Papafitsoros, R. Planqué, W.-P. Vellinga, H. Klinck, T. Denton, P. Bonnet, H. Müller, Lifeclef 2025 teaser: Challenges on species presence prediction and identification, and individual animal identification, Ad...

    work page 2025

  4. [4]

    Picek, M

    L. Picek, M. Šulc, J. Matas, Overview of Fungiclef 2024: Revisiting fungi species recognition beyond 0–1 cost, CLEF 2024 Working Notes CEUR-WS (2024)

    work page 2024

  5. [5]

    Picek, K

    L. Picek, K. Janoušková, V. Cermak, J. Matas, FungiTastic: A multi-modal dataset and benchmark for image categorization, arXiv:2408.13632 (2025)

    arXiv 2025

  6. [6]

    Deitke, C

    M. Deitke, C. Clark, S. Lee, R. Tripathi, Y. Yang, J. S. Park, M. Salehi, N. Muennighoff, K. Lo, L. Soldaini, J. Lu, T. Anderson, E. Bransom, K. Ehsani, H. Ngo, Y. Chen, A. Patel, M. Yatskar, C. Callison-Burch, A. Head, R. Hendrix, F. Bastani, E. VanderBilt, N. Lambert, Y. Chou, A. Chheda, J. Sparks, S. Skjonsberg, M. Schmitz, A. Sarnat, B. Bischoff, P. W...

    arXiv 2024

  7. [7]

    C. Chiu, M. Heil, T. Kim, A. Miyaguchi, Fine-grained classification for poisonous fungi identification with transfer learning, CLEF 2024 Working Notes CEUR-WS (2024)

    work page 2024

  8. [8]

    Z. Liu, H. Hu, Y. Lin, Y. Zhuliang, Z. Xie, Y. Wei, J. Ning, Y. Cao, Z. Zhang, L. Dong, F. Wei, B. Guo, Swin transformer v2: Scaling up capacity and resolution, CVPR 2022, arXiv:2111.09883 (2022)

    arXiv 2022

Show all 28 references
  1. [9]

    S. Wolf, P. H. Thelen, J. Beyerer, Poison-aware open-set fungi classification: Reducing the risk of poisonous confusion, CLEF 2024 Working Notes CEUR-WS (2024)

    2024

  2. [10]

    Goëau, J.-C

    H. Goëau, J.-C. Lombardo, A. Affouard, V. Espitalier, P. Bonnet, A. Joly, PlantCLEF 2024 Pretrained Models on the Flora of Southwestern Europe Based on a Subset of Pl@ntNet Collaborative Images and a ViT Base Patch 14 DINOv2, 2024. URL: https://zenodo.org/records/10848263

    2024

  3. [11]

    Hughes, Demystifying PyTorch’s WeightedRandomSam- pler by example, 2024

    C. Hughes, Demystifying PyTorch’s WeightedRandomSam- pler by example, 2024. URL: https://medium.com/data-science/ demystifying-pytorchs-weightedrandomsampler-by-example-a68aceccb452

    2024

  4. [12]

    Zhang, M

    H. Zhang, M. Cisse, Y. N. Dauphin, D. Lopez-Paz, mixup: Beyond empirical risk minimization, ICLR 2018, arXiv:1710.09412 (2018)

    2018 arXiv

  5. [13]

    URL: http://www

    PACE, Partnership for an Advanced Computing Environment (PACE), 2017. URL: http://www. pace.gatech.edu

    2017

  6. [14]

    URL: https://docs.pytorch.org/docs/stable/generated/torch.nn

    PyTorch, CrossEntropyLoss, 2025. URL: https://docs.pytorch.org/docs/stable/generated/torch.nn. CrossEntropyLoss.html

    2025

  7. [15]

    Poulinakis, Img_Premature_Ending-Detect_Fix.py, 2021

    K. Poulinakis, Img_Premature_Ending-Detect_Fix.py, 2021. URL: https://github. com/Poulinakis-Konstantinos/ML-util-functions/blob/master/scripts/Img_Premature_ Ending-Detect_Fix.py

    2021

  8. [16]

    D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, ICLR 2015, arXiv:1412.6980 (2017)

    2017 arXiv

  9. [17]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V. Vo, M. Szafraniec, V. Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y. Huang, S.-W. Li, I. Misa, M. Rabbat, V. Sharma, G. Synnaeve, H. Xu, H. Jegou, J. Mairal, P. Labatut, A. Jo...

    2024 arXiv

  10. [18]

    H. Bao, L. Dong, S. Piao, F. Wei, Beit: Bert pre-training of image transformers, ICLR 2022, arXiv:2106.08254 (2022)

    2022 arXiv

  11. [19]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, N. Houlsby, An image is worth 16x16 words: Transformers for image recognition at scale, ICLR 2021, arXiv:2010.11929 (2021)

    2021 arXiv

  12. [20]

    Verma, A

    V. Verma, A. Lamb, C. Beckham, A. Najafi, I. Mitliagkas, A. Courville, D. Lopez-Paz, Y. Bengio, Man- ifold mixup: Better representations by interpolating hidden states, ICML 2019, arXiv:1806.05236 (2019)

    2019 arXiv

  13. [21]

    Warner, A

    B. Warner, A. Chaffin, B. Clavié, O. Weller, O. Hallström, S. Taghadouini, A. Gallapher, R. Biswas, F. Ladhak, T. Aarsen, N. Cooper, G. Adams, J. Howard, I. Poli, Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finet...

    2024 arXiv

  14. [22]

    J. Lee, W. Yoon, S. Kim, D. Kim, S. Kim, C. H. So, J. Kang, Biobert: a pre-trained biomedical language representation model for biomedical text mining, Bioinformatics 2019, arXiv:1901.08746 (2019)

    2019 arXiv

  15. [23]

    Z. Chen, V. Badrinarayanan, C.-Y. Lee, A. Rabinovich, Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks, ICML 2018, arXiv:1711.02257 (2018)

    2018 arXiv

  16. [24]

    Stevens, J

    S. Stevens, J. Wu, M. J. Thompson, E. G. Campolongo, C. H. Song, D. E. Carlyn, L. Dong, W. M. Dahdul, C. Stewart, T. Berger-Wolf, W.-L. Chao, Y. Su, Bioclip: A vision foundation model for the tree of life, CVPR 2024, arXiv:2311.18803 (2024)

    2024 arXiv

  17. [25]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, I. Sutskever, Learning transferable visual models from natural language supervision, ICML 2021, arXiv:2103.00020 (2021)

    2021 arXiv

  18. [26]

    G. Team, T. Mesnard, C. Hardin, R. Dadashi, S. Bhupatiraju, S. Pathak, L. Sifre, M. Rivière, M. S. Kale, J. Love, et al., Gemma: Open models based on gemini research and technology, arXiv preprint arXiv:2403.08295 (2024)

    2024 arXiv

  19. [27]

    Abdin, J

    M. Abdin, J. Aneja, H. Awadalla, A. Awadallah, A. A. Awan, N. Bach, A. Bahree, A. Bakhtiari, J. Bao, H. Behl, et al., Phi-3 technical report: A highly capable language model locally on your phone, arXiv preprint arXiv:2404.14219 (2024)

    2024 arXiv

  20. [28]

    Touvron, T

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, et al., Llama: Open and efficient foundation language models, arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

Pith tools

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