kaust-generative-ai/telco-gaia
收藏资源简介:
--- license: cc-by-nc-4.0 pretty_name: Telco-GAIA task_categories: - question-answering language: - en - ar tags: - agents - benchmark - gaia - rag - telco - web-navigation size_categories: - n<1K configs: - config_name: questions data_files: - split: test path: suite/agent_inputs/questions.json --- # Telco-GAIA A GAIA-style benchmark for AI agents operating over a real telecom operator's website snapshot plus a synthetic customer database. **100 tasks** across 7 categories: Pricing, Miscellaneous, Images, Web Archives, PDF, PDF Visual, Database. Agents read `questions.json` + `environment.md`, browse the local website (`:8080`) and query the database API (`:8081`), and produce a GAIA-compatible `submission.json`. ## What's here | File | What | |---|---| | `suite/agent_inputs/questions.json` | The 100 tasks given to the agent (browsable in the viewer above) | | `suite/agent_inputs/environment.md` | Tool/resource description handed to the agent | | `suite/` | Docker harness to serve the website + DB: `Dockerfile`, `docker-compose.yml`, `nginx/` | | `website_snapshot.tar.gz` | Static website snapshot served at `:8080` (LFS) | | `synthetic_db.tar.gz` | Synthetic customer DB served at `:8081` (LFS) | **Ground truth + scorer** are held in a separate **gated** dataset, `kaust-generative-ai/telco-gaia-groundtruth` (answers, `evaluate.py`, `answer_matching.py`), to prevent contamination and eval gaming. Request access there to score submissions. ## Quick start ```bash # 1. get the files huggingface-cli download kaust-generative-ai/telco-gaia --repo-type dataset --local-dir telco-gaia cd telco-gaia # 2. start the environment (website :8080, DB API :8081) # The build auto-extracts website_snapshot.tar.gz / synthetic_db.tar.gz — # no manual unzip needed. cd suite && docker compose up --build -d # 3. run your agent against questions.json + environment.md -> submission.json # 4. evaluate: get access to the gated repo, then use its scorer + answers huggingface-cli download kaust-generative-ai/telco-gaia-groundtruth --repo-type dataset --local-dir gt python gt/evaluate.py --submission submission.json --ground-truth gt/ground_truth.jsonl --output results.json ``` ## License **CC BY-NC 4.0** — Creative Commons Attribution-NonCommercial 4.0 International (https://creativecommons.org/licenses/by-nc/4.0/). Free to use, share, and adapt for **non-commercial research** with attribution. The website snapshot and synthetic DB are a captured copy of a real telecom operator's public site, redistributed for research use only. All site content, page text, trademarks, and logos remain the property of their owner; this release transfers no trademark rights. If you are the rights holder and want content removed or changed, open a discussion here. Please cite Telco-GAIA (EMNLP 2026 Industry Track) — BibTeX to be added.
A GAIA-style benchmark for AI agents operating over a real telecom operators website snapshot plus a synthetic customer database. 100 tasks across 7 categories: Pricing, Miscellaneous, Images, Web Archives, PDF, PDF Visual, Database. Agents read questions.json + environment.md, browse the local website (:8080) and query the database API (:8081), and produce a GAIA-compatible submission.json.




