Model Zoo¶

This page lists OpenSportsLib checkpoints, config-only algorithms, and supporting model components published on Hugging Face. Task checkpoints can be loaded with load_weights(...); config-only algorithms and supporting components use the loading instructions in their sections below.

Scores, class lists, and intended-use statements below come from model cards, published configs, or the library's component configuration. Scores are not computed or verified at documentation-build time. Missing training provenance and scores are marked Not reported. Follow each linked repository for its artifact details; a listing here does not certify end-to-end inference.

Available Models¶

Model	Task	Dataset trained on	Backbone / architecture	Classes / label set	Scores	Hugging Face link	Loading
`OSL-cls-action-mvitv2`	Action / Event Classification	SoccerNet - MVFouls classification subset	MViT v2	Not reported on model card	Accuracy: 0.57 Balanced Accuracy: 0.40 Top-2: 0.78	OpenSportsLab/OSL-cls-action-mvitv2	`myModel.load_weights(weights="OpenSportsLab/OSL-cls-action-mvitv2")`
`OSL-loc-snbas-2023-e2e`	Action Spotting / Localization	SoccerNet - Ball Action Spotting 2023	E2E, DALI backend	PASS, DRIVE	tight mAP: 71.48 loose mAP: 85.62	OpenSportsLab/OSL-loc-snbas-2023-e2e	`myModel.load_weights(weights="OpenSportsLab/OSL-loc-snbas-2023-e2e")`
`OSL-loc-snbas-2025-e2e`	Action Spotting / Localization	SoccerNet - Ball Action Spotting 2025	E2E, DALI backend	PASS, DRIVE, HEADER, HIGH PASS, OUT, CROSS, THROW IN, SHOT, BALL PLAYER BLOCK, PLAYER SUCCESSFUL TACKLE, FREE KICK, GOAL	tight mAP: 47.98 loose mAP: 58.35	OpenSportsLab/OSL-loc-snbas-2025-e2e	`myModel.load_weights(weights="OpenSportsLab/OSL-loc-snbas-2025-e2e")`
`OSL-VQA-XFOUL-XVARS-lora`	Visual Question Answering (VQA)	OSL-XFoul	X-VARS VideoChatGPT + LoRA	Referee-style soccer VQA	Accuracy: 72.24% Balanced Accuracy: 50.00%	OpenSportsLab/OSL-VQA-XFOUL-XVARS-lora	`myModel.load_weights(weights="OpenSportsLab/OSL-VQA-XFOUL-XVARS-lora")`
`OSL-VQA-XFOUL-qwen2.5-7B-VL-lora`	Visual Question Answering (VQA)	OSL-XFoul	Qwen2.5-VL-7B-Instruct + LoRA	Referee-style soccer VQA	Accuracy: 65.77% Balanced Accuracy: 36.59%	OpenSportsLab/OSL-VQA-XFOUL-qwen2.5-7B-VL-lora	`myModel.load_weights(weights="OpenSportsLab/OSL-VQA-XFOUL-qwen2.5-7B-VL-lora")`
`OSL-VQA-XFOUL-qwen3-8B-VL-lora`	Visual Question Answering (VQA)	OSL-XFoul	Qwen3-VL-8B-Instruct + LoRA	Referee-style soccer VQA	Accuracy: 69.96% Balanced Accuracy: 48.33%	OpenSportsLab/OSL-VQA-XFOUL-qwen3-8B-VL-lora	`myModel.load_weights(weights="OpenSportsLab/OSL-VQA-XFOUL-qwen3-8B-VL-lora")`
`skeleton-header-max-recall`	Header spotting	Not applicable: rule-based algorithm	Skeleton maximum-recall rules; CPU, HDF5 tracking inputs	`header`	Not reported	OpenSportsLab/skeleton-header-max-recall	Config only; no weights
`OSL-loc-tennis-public`	Tennis action spotting / localization	Not reported	E2E: `rny002_gsm` + GRU	`no_swing`, `swing` (published config)	Not reported	OpenSportsLab/OSL-loc-tennis-public	Checkpoint and local inputs

Supporting Components¶

These repositories supply components for a configured VQA model. They are not standalone task models to pass to a task wrapper's load_weights(...) method.

Component	Role	Training source	Scores	Hugging Face link	Configuration
`base_model_videoChatGPT`	X-VARS decoder model and tokenizer bundle	Not reported	Not reported	OpenSportsLab/base_model_videoChatGPT	Decoder repository
`trained-clip-vit-large-patch14`	Visual encoder checkpoint for the CLIP ViT-L/14 component	Not reported	Not reported	OpenSportsLab/trained-clip-vit-large-patch14	Visual encoder weights

OSL-cls-action-mvitv2¶

Intended use: video-based soccer action / event classification.

Dataset/training source: SoccerNet - MVFouls classification subset, using video clips.

Reported metrics:

Metric	Score
Accuracy	0.57
Balanced Accuracy	0.40
Top-2	0.78

Hugging Face: OpenSportsLab/OSL-cls-action-mvitv2

myModel.load_weights(weights="OpenSportsLab/OSL-cls-action-mvitv2")

OSL-loc-snbas-2023-e2e¶

Intended use: video-based soccer action spotting / localization.

Dataset/training source: SoccerNet - Ball Action Spotting 2023, using video clips at 224p resolution. The model card reports two classes: PASS and DRIVE.

Reported metrics:

Metric	Score
tight mAP	71.48
loose mAP	85.62

Hugging Face: OpenSportsLab/OSL-loc-snbas-2023-e2e

myModel.load_weights(weights="OpenSportsLab/OSL-loc-snbas-2023-e2e")

OSL-loc-snbas-2025-e2e¶

Intended use: video-based soccer action spotting / localization.

Dataset/training source: SoccerNet - Ball Action Spotting 2025, using video clips at 224p resolution. The model card reports twelve classes: PASS, DRIVE, HEADER, HIGH PASS, OUT, CROSS, THROW IN, SHOT, BALL PLAYER BLOCK, PLAYER SUCCESSFUL TACKLE, FREE KICK, and GOAL.

Reported metrics:

Metric	Score
tight mAP	47.98
loose mAP	58.35

Hugging Face: OpenSportsLab/OSL-loc-snbas-2025-e2e

myModel.load_weights(weights="OpenSportsLab/OSL-loc-snbas-2025-e2e")

OSL-VQA-XFOUL-XVARS-lora¶

Intended use: soccer visual question answering focused on foul analysis and referee-style explanations.

Dataset/training source: OSL-XFoul, using video clips and VQA supervision.

Reported metrics:

Metric	Score
Accuracy	72.24%
Balanced Accuracy	50.00%

Hugging Face: OpenSportsLab/OSL-VQA-XFOUL-XVARS-lora

myModel.load_weights(weights="OpenSportsLab/OSL-VQA-XFOUL-XVARS-lora")

Recommended config:

config="opensportslib/configs/vqa/xvars.yaml"

OSL-VQA-XFOUL-qwen2.5-7B-VL-lora¶

Intended use: native end-to-end soccer visual question answering with Qwen VL.

Dataset/training source: OSL-XFoul, using video clips and VQA supervision.

Reported metrics:

Metric	Score
Accuracy	65.77%
Balanced Accuracy	36.59%

Hugging Face: OpenSportsLab/OSL-VQA-XFOUL-qwen2.5-7B-VL-lora

myModel.load_weights(weights="OpenSportsLab/OSL-VQA-XFOUL-qwen2.5-7B-VL-lora")

Recommended config:

config="opensportslib/configs/vqa/qwen3_vl_native.yaml"

OSL-VQA-XFOUL-qwen3-8B-VL-lora¶

Intended use: native end-to-end soccer visual question answering with Qwen VL.

Dataset/training source: OSL-XFoul, using video clips and VQA supervision.

Reported metrics:

Metric	Score
Accuracy	69.96%
Balanced Accuracy	48.33%

Hugging Face: OpenSportsLab/OSL-VQA-XFOUL-qwen3-8B-VL-lora

myModel.load_weights(weights="OpenSportsLab/OSL-VQA-XFOUL-qwen3-8B-VL-lora")

Recommended config:

config="opensportslib/configs/vqa/qwen3_vl_native.yaml"

skeleton-header-max-recall¶

Hugging Face: OpenSportsLab/skeleton-header-max-recall

This inference-only, rule-based algorithm detects headers from player-joint and ball HDF5 tracking inputs on CPU. The published config selects h5_header_skeleton_max_recall and the label header. It contains no learned checkpoint and does not support training. Reported scores: Not reported.

The repository is gated. Request access on its Hugging Face page and, once approved, authenticate your local Hugging Face client before downloading. Use an OSL manifest referencing your own player-joint and ball HDF5 files; see the HDF5 header spotting guide for the input format.

Download only the config and replace its machine-specific data paths:

from huggingface_hub import hf_hub_download
from opensportslib.apis import Config, LocalizationModel

config_path = hf_hub_download(
    repo_id="OpenSportsLab/skeleton-header-max-recall",
    filename="config.yaml",
)
config = Config.from_file(config_path)
config.update(
    data={
        "data_root": "/path/to/tracking",
        "test_set": "/path/to/header_manifest.json",
    },
    runtime={"device": "cpu"},
)
model = LocalizationModel(config=config, weights=None)
predictions = model.infer(use_wandb=False)
model.save_predictions(output_path="header_predictions.json", predictions=predictions)

Use Config.from_file for this downloaded config: Config.from_pretrained remembers the repository as a weight source, whereas this algorithm has no weights to load.

Server loading: Register this repository as a localization model. The server recognizes its canonical MODEL.metadata.runner and loads it with weights=None. Gated access still requires an approved Hugging Face token. Existing deployments must include the config-only loading fix before using this registration path; see model registration.

OSL-loc-tennis-public¶

Hugging Face: OpenSportsLab/OSL-loc-tennis-public

The repository contains config.yaml and model.pt for E2E tennis spotting. The published config specifies the rny002_gsm backbone, a GRU head, the labels no_swing and swing, and a DALI video loader. Its sampling uses an 11-frame clip at an extraction rate of 2 fps. The current config loader migrates the published legacy config when loading it.

Training source: Not reported. Reported scores: Not reported. The repository has no model card; the architecture and labels above come from its config.

Set local annotations and the root used to resolve their video paths before constructing the model. Config.from_pretrained downloads the config and remembers the repository ID; constructing LocalizationModel then loads the checkpoint.

from opensportslib.apis import Config, LocalizationModel

config = Config.from_pretrained("OpenSportsLab/OSL-loc-tennis-public")
config.update(
    data={
        "data_root": "/path/to/tennis/videos",
        "test_set": "/path/to/tennis/test.json",
    },
    runtime={"device": "cpu"},
)
model = LocalizationModel(config=config)
predictions = model.infer(use_wandb=False)
model.save_predictions(output_path="tennis_predictions.json", predictions=predictions)

The CPU example uses the library's OpenCV fallback. To use the published DALI route, install GPU support and DALI with opensportslib setup --dali, and select device="cuda". Provide an OSL JSON test manifest whose video paths resolve under your chosen data root.

base_model_videoChatGPT¶

Hugging Face: OpenSportsLab/base_model_videoChatGPT

This is the decoder and tokenizer bundle used by the X-VARS VQA backend. The repository contains config.json, sharded safetensors weights, their index, and tokenizer files. Training source: Not reported. Reported scores: Not reported.

The canonical opensportslib/configs/vqa/xvars.yaml already selects this repository. The relevant component setting is:

MODEL:
  components:
    llm_decoder:
      params:
        repo_id: OpenSportsLab/base_model_videoChatGPT

This is a fragment to apply to the X-VARS config, not a complete task config. The decoder works with the configured visual encoder, projector, and task adapter. See X-VARS setup and artifact flow for the workflow and OSL-VQA-XFOUL-XVARS-lora for the task adapter.

trained-clip-vit-large-patch14¶

Hugging Face: OpenSportsLab/trained-clip-vit-large-patch14

This repository supplies model.pth.tar for the CLIP ViT-L/14 visual encoder used by the X-VARS configuration. Training source: Not reported. Reported scores: Not reported.

The canonical opensportslib/configs/vqa/xvars.yaml selects the checkpoint and its corresponding architecture through these component settings:

MODEL:
  components:
    video_encoder:
      load:
        weights_path: OpenSportsLab/trained-clip-vit-large-patch14
      params:
        vision_tower: openai/clip-vit-large-patch14

Apply this fragment to the X-VARS config. Its visual-checkpoint resolver downloads model.pth.tar from the repository; vision_tower identifies the architecture and image processor used with those weights. The encoder supplies video features to the VQA pipeline. See X-VARS setup and artifact flow for feature extraction and index preparation.