Data/add 12 patent datasets

[snowy2002]

No description provided.

[Copilot]

Pull request overview

This PR updates ABDataBench documentation and benchmarking defaults while adding new patent documents to the dataset, and it improves evaluation robustness around sequence/CDRH3 matching plus VLM extraction fallbacks.

Changes:

• Refresh README/README_zh content (features, quick start, dataset links, metrics, and layout).
• Add new patent dataset markdowns (with associated images) under dataset/.
• Update benchmark judge defaults to gzy/gemini-3.1-pro, expand rule-based sequence matching (esp. CDRH3 boundary handling), and harden VLM sequence-image extraction with a chat-completions fallback plus related config/tests.

Reviewed changes

Copilot reviewed 12 out of 896 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
README.md	Rewrites/reshapes English project overview and usage instructions.
README_zh.md	Mirrors README restructuring for Chinese documentation.
dataset/WO2018068182A1/WO2018068182A1.md	Adds a new patent document in markdown with embedded image references.
dataset/WO2018048939A1/WO2018048939A1.md	Adds a new patent document in markdown with embedded image references.
benchmark/scripts/llm_judge.py	Switches default judge model and adds rule-based sequence/CDRH3 boundary normalization logic.
benchmark/run_eval.py	Uses DEFAULT_MODEL from llm_judge and updates example/default model wiring.
benchmark/run_benchmark.sh	Updates default judge model env default.
benchmark/LLM_JUDGE_PROMPT.md	Updates documented default judge model.
agent/tools/sequence_image_tool.py	Adds configurable endpoint selection and fallback from Responses API to chat completions.
agent/tests/test_orchestrator_flow.py	Updates tests for _should_run_vlm_gap_fill becoming instance-based and adds new gating assertions.
agent/tests/test_llm_judge.py	Adds coverage for new default model behavior and new sequence matching rules.
agent/orchestrator.py	Gates sequence-image extractor creation and makes VLM-gap-fill decision depend on instance config state.
agent/config.py	Allows env-based toggling for image extraction and the sequence image tool.
.gitignore	Expands ignores for nested __pycache__ and *.pyc.
.env.example	Updates benchmark model default and clarifies judge credential fallback behavior.

Suppressed comments (1)

agent/tools/sequence_image_tool.py:637

• _auth_header_value returns a literal "******" when the API key is not already prefixed with "Bearer ", which will break authentication for both the Responses and chat-completions calls.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.