Formatting fixes, NEON acknowledgment, and link additions

README.md

@@ -3,6 +3,8 @@ license: mit
 language:
 - en
 library_name: transformers
+base_model:
+- facebook/mask2former-swin-large-ade-semantic
 tags:
 - biology
 - CV
@@ -12,7 +14,8 @@ tags:
 - image-segmentation
 - mask2former
 - entomology
-datasets: imageomics/sentinel-beetles
+datasets:
+- imageomics/sentinel-beetles
 metrics:
 - mean_iou
 model_description: >-
@@ -48,8 +51,8 @@ Mask2Former with a Swin-Large backbone, fine-tuned for pixel-level semantic segm
 
 ### Model Sources
 
-- **Repository:** [https://github.com/Imageomics/BeetleFlow](https://github.com/Imageomics/BeetleFlow)
-- **Paper:** BeetleFlow: An Integrative Deep Learning Pipeline for Beetle Image Processing — *NeurIPS 2025 Workshop for Imageomics: Discovering Biological Knowledge from Images Using AI*
+- **Repository:** [Imageomics/BeetleFlow](https://github.com/Imageomics/BeetleFlow)
+- **Paper:** [BeetleFlow: An Integrative Deep Learning Pipeline for Beetle Image Processing](https://doi.org/10.48550/arXiv.2511.00255) — *NeurIPS 2025 Workshop for Imageomics: Discovering Biological Knowledge from Images Using AI*
 - **Demo:** See [batch inference usage](#how-to-get-started-with-the-model) below
 
 ## Uses
@@ -74,7 +77,7 @@ Mask2Former with a Swin-Large backbone, fine-tuned for pixel-level semantic segm
 
 ## Bias, Risks, and Limitations
 
-- Models are trained on pinned ground-beetle specimens from NEON pitfall-trap imagery; performance may degrade on other beetle families, imaging setups, or specimen preparations.
+- Models are trained on pinned ground-beetle specimens from [NEON](https://www.neonscience.org/) pitfall-trap imagery; performance may degrade on other beetle families, imaging setups, or specimen preparations.
 - The 5-class and 9-class label schemes differ in granularity; choose the checkpoint that matches your annotation protocol.
 - Small or occluded body parts (e.g., legs, antennas) may be harder to segment accurately.
 - Segmentation quality depends on upstream cropping; poorly cropped inputs will reduce mask quality.
@@ -147,7 +150,7 @@ Expect `input/test_images/` and `input/test_masks/` with paired RGB mask images.
 
 ### Training Data
 
-Models are trained on manually annotated beetle part segmentation data from pinned specimen images in the [Beetles as Sentinel Taxa dataset](https://huggingface.co/datasets/imageomics/sentinel-beetles). Dataset details:
+Models are trained on manually annotated beetle part segmentation data from pinned specimen images in the [Beetles as Sentinel Taxa dataset](https://huggingface.co/datasets/imageomics/sentinel-beetles). Details of the subset used for training these models:
 
 **5-class**
 - Parts: head, pronotum, elytra, legs, antennas (+ background)
@@ -159,7 +162,7 @@ Models are trained on manually annotated beetle part segmentation data from pinn
 - 330 labeled beetles with RGB masks
 - Split: 264 training / 66 test images
 
-Training uses `train_images/` and `valid_images/` (with corresponding `train_masks/` and `valid_masks/`) under the data root. See the [dataset README](https://github.com/Imageomics/BeetleFlow/tree/main/data) for mask color conventions.
+Training uses `train_images/` and `valid_images/` (with corresponding `train_masks/` and `valid_masks/`) under the [BeetleFlow GitHub repo](https://github.com/Imageomics/BeetleFlow) data root; this repo includes mask color conventions in the [data README](https://github.com/Imageomics/BeetleFlow/tree/main/data).
 
 ### Training Procedure
 
@@ -169,7 +172,7 @@ Training uses `train_images/` and `valid_images/` (with corresponding `train_mas
 - **Training augmentations** (Albumentations): horizontal flip (p=0.5), random brightness/contrast (p=0.25), rotation (±25°).
 - **Validation augmentations:** resize and normalize only.
 - **Normalization:** ADE20K mean `[123.675, 116.280, 103.530]` and std `[58.395, 57.120, 57.375]` (scaled to [0, 1]).
-- RGB masks are converted to integer class labels using fixed color palettes defined in `config.py`.
+- RGB masks are converted to integer class labels using fixed color palettes defined in [`config.py`](https://github.com/Imageomics/BeetleFlow/blob/main/pipeline/3_segmentation/config.py).
 
 #### Training Hyperparameters
 
@@ -188,7 +191,7 @@ Training was performed on a single NVIDIA A100 GPU (40 GB) with 8 DataLoader wor
 
 ## Evaluation
 
-Evaluation follows `pipeline/3_segmentation/test.py`: predictions are compared against held-out RGB masks at 512×512 resolution using the Hugging Face `evaluate` library.
+Evaluation follows [`pipeline/3_segmentation/test.py`](https://github.com/Imageomics/BeetleFlow/blob/main/pipeline/3_segmentation/test.py): predictions are compared against held-out RGB masks at 512×512 resolution using the Hugging Face `evaluate` library.
 
 ### Testing Data, Factors & Metrics
 
@@ -209,7 +212,7 @@ Held-out test images and RGB masks (`test_images/`, `test_masks/`) from the Beet
 
 ### Results
 
-Evaluated on the held-out test split using `test.py` at 512×512 resolution.
+Evaluated on the held-out test split using [`test.py`](https://github.com/Imageomics/BeetleFlow/blob/main/pipeline/3_segmentation/test.py) at 512×512 resolution.
 
 #### 5-class
 
@@ -257,7 +260,7 @@ Not applicable for this release. Validation segmentation overlays are saved duri
 
 ## Environmental Impact
 
-Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://doi.org/10.48550/arXiv.1910.09700).
+Carbon emissions were estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://doi.org/10.48550/arXiv.1910.09700):
 
 - **Hardware Type:** a single NVIDIA A100 GPU (40 GB)
 - **Hours used:** ~0.5 GPU-hours per model
@@ -333,9 +336,13 @@ If you use these models, please cite BeetleFlow and the underlying dataset.
 
 ## Acknowledgements
 
-This work was supported by the [Imageomics Institute](https://imageomics.org), which is funded by the US National Science Foundation's Harnessing the Data Revolution (HDR) program under [Award #2118240](https://www.nsf.gov/awardsearch/showAward?AWD_ID=2118240) (Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.
+This project was in part conceived at [Funcapalooza](https://github.com/Imageomics/FuncaPalooza-2025).
 
-This work was in part conceived at [Funcapalooza](https://github.com/Imageomics/FuncaPalooza-2025).
+This work was supported by the [Imageomics Institute](https://imageomics.org), which is funded by the US National Science Foundation's Harnessing the Data Revolution (HDR) program under [Award #2118240](https://www.nsf.gov/awardsearch/showAward?AWD_ID=2118240) (Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning).
+
+This material is based in part upon work supported by the [National Ecological Observatory Network (NEON)](https://www.neonscience.org/), a program sponsored by the U.S. National Science Foundation (NSF) and operated under cooperative agreement by Battelle.
+
+Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.
 
 ## Glossary