{"id":50924779,"url":"https://github.com/fcakyon/senben","last_synced_at":"2026-06-16T21:30:46.133Z","repository":{"id":354094381,"uuid":"1206402762","full_name":"fcakyon/senben","owner":"fcakyon","description":"Official repo for CVPR26 paper \"SenBen: Sensitive Scene Graphs for Explainable Content Moderation\"","archived":false,"fork":false,"pushed_at":"2026-04-27T05:23:19.000Z","size":636,"stargazers_count":0,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-04-27T06:25:32.945Z","etag":null,"topics":[],"latest_commit_sha":null,"homepage":"https://senben.kim","language":null,"has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"mit","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/fcakyon.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-04-09T22:03:41.000Z","updated_at":"2026-04-27T05:17:28.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/fcakyon/senben","commit_stats":null,"previous_names":["fcakyon/senben"],"tags_count":1,"template":false,"template_full_name":null,"purl":"pkg:github/fcakyon/senben","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/fcakyon%2Fsenben","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/fcakyon%2Fsenben/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/fcakyon%2Fsenben/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/fcakyon%2Fsenben/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/fcakyon","download_url":"https://codeload.github.com/fcakyon/senben/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/fcakyon%2Fsenben/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":34425020,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-06-16T02:00:06.860Z","response_time":126,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":[],"created_at":"2026-06-16T21:30:45.468Z","updated_at":"2026-06-16T21:30:46.125Z","avatar_url":"https://github.com/fcakyon.png","language":null,"funding_links":[],"categories":[],"sub_categories":[],"readme":"# 🎬 SenBen: Sensitive Scene Graphs for Explainable Content Moderation\n\n[![arXiv](https://img.shields.io/badge/arXiv-2604.08819-b31b1b.svg)](https://arxiv.org/abs/2604.08819)\n[![HF Papers](https://img.shields.io/badge/🤗-Papers-yellow)](https://huggingface.co/papers/2604.08819)\n[![Dataset](https://img.shields.io/badge/🤗-Dataset-orange)](https://huggingface.co/datasets/fcakyon/senben)\n[![Models](https://img.shields.io/badge/🤗-Models-orange)](https://huggingface.co/fcakyon/senben-q2l-bal)\n[![Project page](https://img.shields.io/badge/site-senben.kim-blueviolet)](https://senben.kim)\n[![License](https://img.shields.io/badge/license-MIT-green)](LICENSE)\n\nCVPR Workshops 2026 paper by [Fatih Cagatay Akyon](https://scholar.google.com/citations?user=RHGyDE0AAAAJ\u0026hl=en) and [Alptekin Temizel](https://blog.metu.edu.tr/atemizel/). 13,999 movie frames annotated with grounded scene graphs for evaluating sensitive content understanding in vision language models, plus a 241M Florence-2 student that beats every evaluated VLM except Gemini on grounded metrics.\n\n## Resources\n\n|                                 |                                                                                                                                                            |\n| ------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- |\n| Paper                           | [arXiv 2604.08819](https://arxiv.org/abs/2604.08819) · [PDF](https://arxiv.org/pdf/2604.08819.pdf) · [HF Papers](https://huggingface.co/papers/2604.08819) |\n| Project page                    | https://senben.kim                                                                                                                                         |\n| Dataset (gated, v1.0)           | https://huggingface.co/datasets/fcakyon/senben                                                                                                             |\n| Model: Q2L-bal (high tag F1)    | https://huggingface.co/fcakyon/senben-q2l-bal                                                                                                              |\n| Model: Q2L-agg (high SenBen-F1) | https://huggingface.co/fcakyon/senben-q2l-agg                                                                                                              |\n| Predictions release             | https://github.com/fcakyon/senben/releases/tag/v1.0-results                                                                                                |\n\n## Predictions and eval metrics (v1.0-results)\n\nPer-frame predictions and machine-readable metrics for every model evaluated in the paper on the SenBen v1.0 test split (2,000 frames). Each tarball extracts to a directory with one JSON per frame plus `_meta.json`, `_summary.json`, and (where applicable) `_tag_metrics.json`.\n\n### Our models (paper Tables 2 + 3 + 4)\n\n| Tarball                                                                                                                   | Params | SenBen-Recall | SenBen-F1 | Tag F1 | Object Recall | Caption Similarity |\n| ------------------------------------------------------------------------------------------------------------------------- | -----: | ------------: | --------: | -----: | ------------: | -----------------: |\n| [`senben_q2l_agg__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/senben_q2l_agg__v1.tar.gz) |   241M |          .449 |      .431 |   .457 |          .409 |               .772 |\n| [`senben_q2l_bal__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/senben_q2l_bal__v1.tar.gz) |   241M |          .413 |      .428 |   .594 |          .420 |               .771 |\n\n### Frontier vision language models (paper Table 2)\n\nSorted by SenBen-F1 descending.\n\n| Tarball                                                                                                                                       | Model                      |      Params | SenBen-Recall | SenBen-F1 | Tag F1 | Object Recall | Caption Similarity |\n| --------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------- | ----------: | ------------: | --------: | -----: | ------------: | -----------------: |\n| [`baseline_gemini3pro__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_gemini3pro__v1.tar.gz)           | Gemini 3 Pro (low reas.)   | proprietary |          .652 |      .647 |   .806 |          .295 |               .642 |\n| [`baseline_gemini3flash__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_gemini3flash__v1.tar.gz)       | Gemini 3 Flash (low reas.) | proprietary |          .593 |      .583 |   .784 |          .271 |               .654 |\n| [`baseline_claude_opus46__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_claude_opus46__v1.tar.gz)     | Claude Opus 4.6            | proprietary |          .327 |      .404 |   .658 |          .082 |               .598 |\n| [`baseline_glm46v_flash__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_glm46v_flash__v1.tar.gz)       | GLM-4.6V (reas.)           |       10.3B |          .291 |      .364 |   .492 |          .123 |               .563 |\n| [`baseline_gpt52_thinking__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_gpt52_thinking__v1.tar.gz)   | GPT-5.2 (med. reas.)       | proprietary |          .319 |      .362 |   .608 |          .072 |               .616 |\n| [`baseline_qwen3vl_8b__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_qwen3vl_8b__v1.tar.gz)           | Qwen3-VL-8B                |        8.3B |          .286 |      .340 |   .469 |          .104 |               .548 |\n| [`baseline_claude_sonnet46__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_claude_sonnet46__v1.tar.gz) | Claude Sonnet 4.6          | proprietary |          .277 |      .339 |   .643 |          .034 |               .590 |\n| [`baseline_gpt5_mini__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_gpt5_mini__v1.tar.gz)             | GPT-5-mini (med. reas.)    | proprietary |          .285 |      .330 |   .659 |          .040 |               .605 |\n| [`baseline_gpt52__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_gpt52__v1.tar.gz)                     | GPT-5.2                    | proprietary |          .247 |      .304 |   .550 |          .052 |               .583 |\n\n### Commercial safety APIs and classifiers (paper Table 3)\n\nSorted by Tag F1 descending. F1\u003csup\u003esafe\u003c/sup\u003e is binary safe vs unsafe F1 over the full taxonomy.\n\n| Tarball                                                                                                                                                 | Model                |      Params | Tags supported | Tag F1 | F1\u003csup\u003esafe\u003c/sup\u003e |\n| ------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------- | ----------: | -------------: | -----: | ----------------: |\n| [`baseline_azure_content_safety__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_azure_content_safety__v1.tar.gz) | Azure Content Safety | proprietary |         5 / 16 |   .430 |              .504 |\n| [`baseline_openai_moderation__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_openai_moderation__v1.tar.gz)       | OpenAI Moderation    | proprietary |         6 / 16 |   .411 |              .664 |\n| [`baseline_llavaguard__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_llavaguard__v1.tar.gz)                     | LlavaGuard 1.2       |        7.0B |         6 / 16 |   .384 |              .583 |\n| [`baseline_google_safesearch__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_google_safesearch__v1.tar.gz)       | Google SafeSearch    | proprietary |         8 / 16 |   .341 |              .476 |\n| [`baseline_sd_safety_checker__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_sd_safety_checker__v1.tar.gz)       | SD Safety Checker    |        304M |         2 / 16 |   .333 |              .472 |\n| [`baseline_nudenet_detector__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_nudenet_detector__v1.tar.gz)         | NudeNet Detector     |       25.9M |         1 / 16 |   .238 |              .238 |\n| [`baseline_laion_safety_checker__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_laion_safety_checker__v1.tar.gz) | LAION Safety Checker |        1.0B |         2 / 16 |   .225 |              .357 |\n| [`baseline_nudenet_classifier__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_nudenet_classifier__v1.tar.gz)     | NudeNet Classifier   |        8.5M |         1 / 16 |   .117 |              .117 |\n| [`baseline_shieldgemma2__v1.tar.gz`](https://github.com/fcakyon/senben/releases/download/v1.0-results/baseline_shieldgemma2__v1.tar.gz)                 | ShieldGemma 2        |        4.0B |         4 / 16 |   .089 |              .161 |\n\n### Machine-readable metrics\n\n[`senben_eval_metrics__v1.json`](https://github.com/fcakyon/senben/releases/download/v1.0-results/senben_eval_metrics__v1.json) contains paper Tables 1 to 5 (system ablation, leave-one-out per-category, VLM baselines, safety classifiers, efficiency) in a single JSON.\n\n## Code\n\nInference, evaluation, training, and 18 baseline runners are in progress and will land in a follow-up commit. Watch the repo for updates.\n\n## License\n\nCode and trained model weights: MIT (see [`LICENSE`](LICENSE)). The SenBen dataset is gated research only, see the [dataset card](https://huggingface.co/datasets/fcakyon/senben).\n\n## Cite\n\n```bibtex\n@inproceedings{akyon2026senben,\n  title     = {SenBen: Sensitive Scene Graphs for Explainable Content Moderation},\n  author    = {Akyon, Fatih Cagatay and Temizel, Alptekin},\n  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)},\n  year      = {2026},\n  url       = {https://arxiv.org/abs/2604.08819}\n}\n```\n\n`fatih.akyon@metu.edu.tr`\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Ffcakyon%2Fsenben","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Ffcakyon%2Fsenben","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Ffcakyon%2Fsenben/lists"}