Hugging Face's logo
Activity Feed Request to join this org

AI & ML interests

We develop infrastructure for the evaluation of generated text.
albertvillanovaย 
posted an update 6 days ago
view post
Post
186
Day-0 Mistral Large 4 support in TRL ๐Ÿš€

- LoRA SFT of the 1T-A52B model on one 8xB300 node, on top of the FP8 checkpoint
- Experts split 8 ways with expert parallelism, everything else sharded with FSDP, all set at load time
- Assistant-only loss out of the box with its training chat template
- ~32.7k tokens per step in ~8 s

Script + Slurm launcher ๐Ÿ‘‡
https://github.com/huggingface/trl/tree/main/examples/sft_mistral_large_4
https://huggingface.co/posts/sergiopaniego/408182860646834
Parveshiiiiย 
posted an update 7 days ago
view post
Post
2,770
Most AI image detectors work fine in the lab and fail on social media.

The reason is simple. Real photos get compressed, resized and re-uploaded again and again, and detectors trained on clean images read that damage as a sign of fakery. A real photo that has passed through a few platforms often gets flagged as AI-generated.

AIRealNet-2 is built to fix that. It is the upgrade to our original Modotte/AIRealNet, trained on images collected from platforms like Instagram and X, so it has seen the kinds of quality loss and compression real images go through online.

The training set has 98k images, split 50/50 between AI-generated and real, so the model has no class bias to lean on. Personal photos were excluded by design, including AI-edited ones.

The architecture is unchanged: a SwinV2 Tiny backbone with a binary classification head, about 0.2B parameters, working at 256x256 resolution. It stays fast and easy to deploy for research or production.

Model: Modotte/AIRealNet-2
  • 4 replies
ยท
Parveshiiiiย 
posted an update 18 days ago
view post
Post
199
Give your LLM a window to the web ๐Ÿ”Ž

I built Seekify, a Python tool that turns web search into clean, LLM-ready input for free.

Collecting sources for an LLM workflow usually means juggling search APIs, scraping pages, and cleaning up messy text. Seekify handles that in one place:

- No search API key needed
- Structured results with titles, URLs, and snippets
- Optional page text extraction so your model gets real content, not just links

Search the web, collect sources, and feed them straight into your pipeline, so you spend less time gathering data and more time building.

It's open source, and I'd love your feedback, ideas, and contributions.

๐Ÿ‘‰ https://github.com/Parveshiiii/Seekify

If it's useful, a โญ on GitHub helps a lot!
Parveshiiiiย 
posted an update 26 days ago
view post
Post
165
Most NSFW classifiers break the second an image touches the internet.

They look great on pristine benchmarks, but in the wild, every social platform aggressively recompresses, downsamples, and degrades images.
The moment JPEG or WebP compression artifacts show up, confidence collapses and false positives spike.

SafeScan was built to survive actual platform pipelines.
Trained on 34,000 images under almost every major social media compression profile using a Vision Transformer backbone (google/vit-base-patch16-224). Instead of blunt binary filtering, it breaks decisions down across 5 clear categories:

โ€ข safe
โ€ข drawing
โ€ข sexy
โ€ข hentai
โ€ข porn

The result is a moderation model that actually generalizes to real-world internet feeds instead of fragile, uncompressed datasets.
Open-weight and available on Hugging Face:

Model: Parveshiiii/SafeScan
Parveshiiiiย 
posted an update 29 days ago
view post
Post
3,854
Most deepfake audio detectors are quietly cheating.

They donโ€™t really listen to the speech โ€” they just look at how long the embedding vector is. Once they figure that out, accuracy looks great on paper and falls apart in the wild.

AIRealNet-Audio was built to stop that shortcut.
It forces every feature onto the unit hypersphere (twice) so the model can only use direction, not magnitude. Trained on speech from 100+ different TTS and voice-cloning systems, plus real human recordings under heavy compression and noise.

The result is a detector that actually has to learn the artifacts instead of gaming the feature space.

Model: Modotte/AIRealNet-Audio
dronefreakย 
posted an update about 1 month ago
view post
Post
367
๐Ÿš€ Excited to open-source the Brackish Underwater Object Detection Model Zoo on Hugging Face.

This release includes:

- ๐Ÿค– 8 YOLOv8, YOLOv11, and YOLOv26 object detection models trained on Brackish, spanning nano through x-large variants (RF-DETR variants to follow once I have the compute for them).
- ๐ŸŸ Benchmarked on Brackish's underwater marine-animal detection task โ€” crab, fish, jellyfish, shrimp, small_fish, and starfish, captured 9 meters below the surface under naturally varying visibility conditions.
- ๐Ÿ“Š Detailed model cards with mAP/precision/recall, per-class breakdowns, PR/F1 curves and confusion matrices, qualitative detection showcases, and full training configurations for reproducibility.

Headline numbers:
- ๐Ÿ† Best mAP@50: 99.3% (YOLOv8s), 85.65% mAP@50:95, 99.36% precision.
- โšก Best efficiency tradeoff: YOLOv26n hits 98.95% mAP@50 at just 6.1 GFLOPs (2.6M params) โ€” within 0.35 points of the top model while using ~4.7x fewer FLOPs (28.6 GFLOPs). The whole zoo actually clusters tightly here (98.74โ€“99.3% mAP@50 across all 8 models), so there's little accuracy left on the table by going small on this dataset.

If you're working on underwater perception, marine biology/ecology monitoring, or robotics in low-visibility environments, I hope these resources are useful.

๐Ÿ“ฆ Dataset:
dronefreak/Brackish

๐Ÿค– Model Collection: dronefreak/brackish-underwater-object-detection-model-zoo-6a7c153687b2821da0c6591d

Feedback, bug reports, and contributions are always welcome.
dronefreakย 
posted an update about 1 month ago
view post
Post
136
๐Ÿš€ Open-sourcing the KITTI Object Detection Model Zoo on Hugging Face.

- ๐Ÿค– 12 models: YOLOv8, YOLOv9, YOLO11 and YOLO26, from nano/tiny up to x-large.
- ๐Ÿš— Street-scene detection: cars, cyclists, pedestrians, vans, trucks and more, in KITTI's ultra-wide frames.
- ๐Ÿ“Š Model cards with metrics, per-class results, curves, showcases, a demo video and full configs.

Headline numbers:
- ๐Ÿ† Best mAP@50: 43.73% (YOLO26x). Best mAP@50:95: 26.54% (YOLO26s).
- โšก YOLO26n gets 42.54% mAP@50 at just 6.1 GFLOPs, within 1.2 points of YOLO26x at ~34x fewer FLOPs.

Trained and evaluated with DetectionBench: https://github.com/dronefreak/DetectionBench

Dataset credit: Andreas Geiger, Philip Lenz and Raquel Urtasun (CVPR 2012). This is an unofficial YOLO-ready reformat (CC BY-NC-SA 3.0). Metrics are on the validation split, since KITTI has no public test labels.

๐Ÿ“ฆ Dataset: dronefreak/KITTI
๐Ÿค– Collection: dronefreak/kitti-object-detection-model-zoo-6ab4cb94fb08dccf3bde2e8b

Feedback and contributions welcome.
dronefreakย 
posted an update about 1 month ago
view post
Post
124
๐Ÿš€ Open-sourcing the BDD100K Object Detection Model Zoo on Hugging Face.

- ๐Ÿค– 10 models: YOLOv8/v9/v10/YOLO11/YOLO26 (n/s) and RF-DETR Nano.
- ๐Ÿš— Driving-scene detection: 10 classes including traffic lights and signs, across varied weather and lighting.
- ๐Ÿ“Š Model cards with metrics, per-class results, curves, showcases, dashcam demo videos and full configs.

Headline numbers:
- ๐Ÿ† Best mAP@50: 58.76% (YOLO26s), 33.86% mAP@50:95.
- โšก YOLO26n gets 52.25% mAP@50 at 6.1 GFLOPs, ~3.7x fewer than YOLO26s.

Trained and evaluated with DetectionBench: https://github.com/dronefreak/DetectionBench

Dataset credit: Fisher Yu et al. (UC Berkeley, CVPR 2020). It has a non-commercial license, so it is not mirrored; get it from https://www.bdd100k.com/. The "test" split here is BDD100K's official validation set.

๐Ÿค– Collection: dronefreak/bdd100k-object-detection-model-zoo-6aafc46f2f6c5e4d8676d894

Feedback and contributions welcome.
dronefreakย 
posted an update about 1 month ago
view post
Post
60
๐Ÿš€ Open-sourcing the UAVDT Object Detection Model Zoo on Hugging Face.

- ๐Ÿค– 18 models: YOLOv8/v9/v10/YOLO11/YOLO26 (nโ€“m) and RF-DETR Nano/Small/Medium.
- ๐Ÿš Drone traffic surveillance: cars, trucks and buses, mostly tiny objects (median box ~0.14% of the image).
- ๐Ÿ“Š Model cards with metrics, per-class results, curves, showcases, demo videos and full configs.

Headline numbers:
- ๐Ÿ† Best mAP@50: 33.43% (YOLOv26m). Best mAP@50:95: 20.54% (RF-DETR Medium).
- โšก YOLOv26s gets 32.98% mAP@50 at 22.8 GFLOPs, ~3.3x fewer than YOLOv26m.

Trained and evaluated with DetectionBench: https://github.com/dronefreak/DetectionBench

Dataset credit: Dawei Du et al. (ECCV 2018). It is research-only, so I don't mirror it; the dataset repo is a guide to the official source.

๐Ÿ“ฆ Dataset: dronefreak/UAVDT
๐Ÿค– Collection: dronefreak/uavdt-object-detection-model-zoo-6aadb3673084702d3ed45ff0

Feedback and contributions welcome.
dronefreakย 
posted an update 3 months ago
view post
Post
442
๐ŸŒง๏ธโ„๏ธ Free demo: remove rain, raindrops, or snow from a photo with a single model

I put together an unofficial demo for **Histoformer** (ECCV 2024, arXiv: 2407.10172), a 16.6M-parameter transformer that handles three different weather degradations, rain streaks, adherent raindrops, and snow, in one unified model. It uses a "histogram self-attention" mechanism that groups pixels by degradation intensity instead of spatial position, which is a
neat way to sidestep the usual spatial-window tradeoffs in restoration transformers.

Try it here, free on ZeroGPU: dronefreak/histoformer-weather-restoration

Upload a photo and get a before/after slider. Two checkpoints available: one tuned for real-world photos, one for the paper's synthetic benchmarks.

Also put together a cleaner, easy-to-use model card with a copy-pasteable Quickstart if you'd rather run it yourself: dronefreak/Histoformer

This is an unofficial demo/mirror, not affiliated with the original authors. All credit for the actual research goes to Shangquan Sun, Wenqi Ren, Xinwei Gao, Rui Wang, and Xiaochun Cao (@sunsean ). Official repo: https://github.com/sunshangquan/Histoformer. Weights are MIT-licensed.

Reported numbers from the paper: 32.1 PSNR on rain+fog (Outdoor-Rain), 33.1 on raindrops, 37.4 / 32.2 on light/heavy snow (Snow100K-S/L).
  • 2 replies
ยท
ndaheimย 
authored a paper 3 months ago
dronefreakย 
posted an update 3 months ago
view post
Post
2,144
๐Ÿš€ Excited to open-source the SeaDronesSee Object Detection Model Zoo on Hugging Face.

This release includes:

- ๐Ÿค– YOLOv8, YOLOv11, YOLOv26 and RF-DETR object detection models trained on SeaDronesSee, spanning nano through x-large YOLO variants plus RF-DETR Nano/Small/Medium.
- ๐ŸŒŠ Benchmarked on SeaDronesSee's maritime search-and-rescue setting โ€” swimmers, boats, jet skis, life-saving appliances and buoys captured by UAVs over open water, at varying altitudes and non-uniform image resolutions (1080p up to 4K+).
- ๐Ÿ“Š Detailed model cards with mAP/precision/recall, per-class breakdowns, PR/F1 curves and confusion matrices (YOLO), qualitative detection showcases, and full training configurations for reproducibility.

Headline numbers:
- ๐Ÿ† Best mAP@50: 83.47% (RF-DETR Medium), 47.49% mAP@50:95, 87.01% precision.
- โšก Best efficiency tradeoff: YOLOv26s hits 80.14% mAP@50 at just 22.8 GFLOPs (10.0M params) โ€” within ~3 points of the top RF-DETR variant, while actually beating YOLOv11x's 74.82% mAP@50 using ~8.6x fewer FLOPs (196.0 GFLOPs).

The goal is to make benchmarking and experimenting with maritime UAV perception easier by providing ready-to-use pretrained checkpoints, all trained and evaluated under one shared pipeline (DetectionBench: https://github.com/dronefreak/DetectionBench).

Full credit for the underlying dataset goes to Leon Amadeus Varga, Benjamin Kiefer, Martin Messmer, and Andreas Zell (University of Tรผbingen, WACV 2022) โ€” this release is an unofficial, YOLO-ready reformatting of their work (CC0-licensed), not a new dataset.

If you're working on maritime search-and-rescue, UAV perception, autonomous drones, or real-time object detection, I hope these resources are useful.

๐Ÿ“ฆ Dataset:
dronefreak/SeaDronesSee

๐Ÿค– Model Collection: dronefreak/seadronessee-object-detection-model-zoo-6a7b030a25797e5dd2d70123

Feedback, bug reports, and contributions are always welcome.
  • 2 replies
ยท
dronefreakย 
posted an update 3 months ago
view post
Post
1,920
๐Ÿš€ Excited to open-source the GWHD Wheat Head Detection Model Zoo on Hugging Face.

This release includes:

- ๐Ÿค– YOLOv8, YOLOv11, YOLOv26 and RF-DETR object detection models trained on GWHD (Global Wheat Head Dataset), spanning nano through x-large variants across both architecture families.
- ๐ŸŒพ Benchmarked on GWHD's dense, single-class wheat-head detection task โ€” ~45 annotated heads per image on average, captured across multiple countries, genotypes, and growth stages, a genuinely hard small/dense-object setting.
- ๐Ÿ“Š Detailed model cards with mAP/precision/recall, per-class breakdowns, PR/F1 curves and confusion matrices (YOLO), qualitative detection showcases, and full training configurations for reproducibility.

Headline numbers:
- ๐Ÿ† Best mAP@50: 74.25% (YOLOv11x), 34.92% mAP@50:95, 83.37% precision.
- โšก Best efficiency tradeoff: YOLOv26s hits 70.49% mAP@50 at just 22.8 GFLOPs (10.0M params) โ€” within ~4 points of the top YOewer FLOPs (196.0 GFLOPs).

The goal is to make benchmarking and experimenting with agricultural computer vision easier by providing ready-to-use pretrained checkpoints, all trained and evaluated under one shared pipeline (DetectionBench: https://github.com/dronefreak/DetectionBench).
Full credit for the underlying dataset goes to Etienne David, Mario Serouart, Simon Madec, and the Global Wheat Head Detection 2020/2021) โ€” this release is anunofficial, YOLO-ready reformatting of their work, not a new dataset.

If you're working on precision at detection, or just want areproducible detector benchmark, I hope these resources are useful.

๐Ÿ“ฆ Dataset:
dronefreak/GWHD

๐Ÿค– Model Collection: https://huggingface.co/collections/dronefreak/gwhd-wheat-head-detection-model-zoo-6a7aea28b5431918cc46cec1

Feedback, bug reports, and contributions are always welcome.
  • 3 replies
ยท
dronefreakย 
posted an update 3 months ago
view post
Post
938
๐Ÿš€ Excited to open-source the VDD Semantic Segmentation Model Zoo on Hugging Face.

This release includes:

- ๐Ÿค– CABiNet and YOLO26 semantic segmentation models trained on VDD (Varied Drone Dataset), spanning Nano through XLarge YOLO26 variants plus a CABiNet (MobileNetV3-Large) baseline.
- ๐ŸŒ Benchmarked on VDD's varied altitudes, viewpoints, and scenes (urban, rural, natural) โ€” a more diverse and challenging setting than single-flight UAV footage.
- ๐Ÿ“Š Detailed model cards with evaluation metrics, per-class IoU, confusion matrices, qualitative RGB / Ground-Truth / Prediction comparisons, and training configurations for reproducibility.

Headline numbers:
- ๐Ÿ† Best mIoU: 78.83% (YOLO26x-sem)
- โšก Best efficiency tradeoff: CABiNet-Large hits 77.76% mIoU at just 54.8 GFLOPs โ€” within 1-2 points of the top YOLO26 variantO26x's 430.9 GFLOPs)

The goal is to make benchmarking and experimenting with aerial semantic segmentation easier by providing ready-to-use pretraineat, all trained and evaluatedunder one shared pipeline.

If you're working on UAV perception, autonomous drones, robotics, remote sensing, or real-time semantic segmentation, I hope these resources are useful.

๐Ÿ“ฆ Dataset: RussRobin/VDD

๐Ÿค– Model Collection: https://huggingface.co/collections/dronefreak/vdd-semantic-segmentation-model-zoo

Feedback, bug reports, and contributions are always welcome.
dronefreakย 
authored 4 papers 4 months ago
dronefreakย 
posted an update 4 months ago
view post
Post
4,569
๐Ÿš€ Excited to open-source the **UAVid Semantic Segmentation Model Zoo** on Hugging Face.

This release includes:

* ๐Ÿ“ฆ A **YOLO-compatible mirror** of the UAVid semantic segmentation dataset, preserving the original train/val/test splits while reorganizing the directory structure for plug-and-play use with modern training pipelines.
* ๐Ÿค– Multiple **YOLO26 semantic segmentation models** trained on UAVid, spanning Nano through Medium variants.
* ๐Ÿ“Š Detailed model cards with evaluation metrics, per-class IoU, confusion matrices, qualitative results, and training configurations for reproducibility.

The goal is to make benchmarking and experimenting with aerial semantic segmentation easier by providing ready-to-use datasets and pretrained models in a consistent format.

If you're working on UAV perception, autonomous drones, robotics, remote sensing, or real-time semantic segmentation, I hope these resources are useful.

**๐Ÿ“ฆ Dataset:** dronefreak/UAVid-2020

**๐Ÿค– Model Collection:** https://huggingface.co/collections/dronefreak/uavid-semantic-segmentation-model-zoo

Feedback, bug reports, and contributions are always welcome.
albertvillanovaย 
posted an update 4 months ago
view post
Post
4,594
๐ŸŽ‰ KTO is now part of the stable TRL API

As of Promote KTO to stable API, KTOTrainer and KTOConfig have graduated from trl.experimental to the stable trl API. https://github.com/huggingface/trl/pull/6175

This one closes out a long road. Over the past 6+ months, the "Align KTO with DPO" effort landed ~90 PRs methodically bringing KTO up to the standard we hold for stable trainers, one carefully-scoped change at a time:
- Feature parity with DPO: full VLM support (incl. multi-image), sync_ref_model, PEFT + Liger, ZeRO-3 + PEFT dtype fix, pad_to_multiple_of, activation offloading, IterableDataset and dict eval_dataset, remove_unused_columns, and reference-logprob precomputation at init.
- Consistency with DPO: aligned method order and signatures, tokenization, _prepare_dataset, PEFT handling, ref-model preparation for distributed training, and config layout โ€” plus a new DataCollatorForKTO and output format. Metrics moved into _compute_loss and simplified to direct averages via the shared _metrics attribute.
- Removing legacy baggage: dropped encoder-decoder support, BOS/EOS handling, null_ref_context, generate_during_eval, model_init, preprocess_logits_for_metrics, model/ref adapter names, and several dead config knobs.
- Coverage: a full test suite mirroring DPO, text collator tests, VLM tests, and slow tests.
- The promotion itself: the experimental โ†’ stable move (#6175) and shim cleanup (#6287), handled so downstream users get a clean deprecation path.

Honestly, this has been one of the more complex tasks I've taken on since joining the team, not because any single change was hard, but because it demanded sustained consistency across a ~2,000-line trainer, with every branch, comment, and edge case kept in lockstep with DPO.

Huge thanks to everyone who reviewed along the way (especially @qgallouedec ), the incremental review cadence is exactly what kept this maintainable.

KTO now sits on equal footing with our other flagship trainers. ๐Ÿš€
  • 2 replies
ยท