RAG-EDA

Structures of the directory:

• benchmark
  • ORD-QA.jsonl (ORD-QA benchmark containing 90 question-reference-answer triplets)
  • openroad_documentation.json (OpenROAD documentation chunks)
• training_dataset
  • generator_dataset
    • eda_corpus_pretrain.jsonl (corpus for generator pre-train)
    • QA_finetuning_v1v2amend1.jsonl (corpus for generator fine-tuning)
  • reranker_dataset
    • openroad_pos_neg_backup_data.jsonl (corpus for reranker finetuning)
  • text-embedding-model_dataset
    • embedding_contrastive_corpus.csv (corpus for text embedding model finetuning)

One example of QA in ORD-QA

{
    "id": 67,
    "type": "functionality",
    "question": " Once the design is routed, how can I estimate the parasitics?\n",
    "ref_num": 1,
    "reference": [
        "global_routing_12"
    ],
    "reference_content": [
        "id:global_routing_12\n### Estimate Global Routing Parasitics\n\nTo estimate RC parasitics based on global route results, use the `-global_routing`\noption of the `estimate_parasitics` command.\n\n```{note}\nTo see the function definition for `estimate_parasitics`, refer to \n[Resizer docs](../rsz/README.md#estimate-parasitics).\n```\n\n```tcl\nestimate_parasitics -global_routing\n```\n\n## Commands\n\n```{note}\n- Parameters in square brackets `[-param param]` are optional.\n- Parameters without square brackets `-param2 param2` are required.\n```\n\n"
    ],
    "answer": " You can estimate the parasitics after global routing by running:\\n\\n```tcl\\nestimate_parasitics -global_routing\\n```\\n\\n\n"
}

• id: the id of the QA.
• type: the question type. There are three question types in total, namely, functionality, vlsi_flow and gui & installation & test.
• question.
• ref_num: number of document(s) used for answering the question.
• reference: the id(s) of the relevant document(s).
• reference_content: the content of the relevant document(s).
• answer.