vgandhi13/grasp-retrieval-data-and-index

GRASP — Training Data and Retrieval Indices

Training/eval parquets and prebuilt retrieval indices for GRASP, an agentic-RAG setup that fine-tunes Qwen2.5-3B/7B-Instruct with GRPO/PPO on HotpotQA distractor using three tools: semantic_search, keyword_search, and read_chunk.

38.3 GB across 28 files.

Layout

Final/
├── 3b_main/          train.parquet (90,447 rows), val_test_256.parquet
├── 7B_main/          train_17344.parquet
├── ablation/         train_5500_<variant>, val_test_256_<variant>
│                     variants: both, no_keyword, no_semantic, no_rc_paragraph
├── Test_Set/         hotpotqa_test_500_<variant>, musique_test_500, twowiki_test_500
└── index/
    ├── train/
    │   ├── hotpotqa_sentence/    sentence_index.pkl (16 GB, unsplit)
    │   │                         sentence_index_bm25.pkl (1.8 GB)
    │   │                         sentence_index_semantic.pkl (14.9 GB)
    │   └── hotpotqa_paragraph/   paragraph_index_{bm25,semantic}.pkl
    └── test/
        ├── hotpotqa_sentence/    sentence_index.pkl (combined BM25 + semantic)
        ├── hotpotqa_paragraph/   paragraph_index.pkl
        ├── musique/              sentence_index_{bm25,semantic}.pkl
        └── twowiki/              sentence_index_{bm25,semantic}.pkl

The train HotpotQA sentence index ships both the unsplit sentence_index.pkl and the split BM25/semantic pair derived from it. The split pair is what the retrieval servers load (BM25 on CPU workers, semantic on GPU workers); the unsplit file is kept for reference. If you only need to run retrieval, skip the 16 GB unsplit file — it roughly halves the download.

Index schema

Each pickle is a dict:

Key	Contents
sentences	list of sentence (or paragraph) strings
embeddings	float32, L2-normalized, dim 1024 — Qwen3-Embedding-0.6B
sentence_to_chunk	index → chunk id (identity for paragraph indices)
chunks	dict id → {id, title, text}
bm25_index	rank_bm25.BM25Okapi
bm25_tokenized	tokenized corpus
model_name	embedding model id

Split files carry only the keys their server needs: *_bm25.pkl has the BM25 side, *_semantic.pkl has embeddings + model_name.

Paragraph indices use the same schema, but each sentences entry is a full paragraph and sentence_to_chunk is the identity map.