| --- |
| pretty_name: TRQA Experiments |
| tags: |
| - retrieval |
| - information-retrieval |
| - faiss |
| - lucene |
| --- |
| |
| # TRQA Experiments |
|
|
| Prebuilt retrieval indices and experiment results accompanying the |
| [**TRQA dataset**](https://dl.acm.org/doi/pdf/10.1145/3805712.3808629). |
|
|
| This repo holds only artifacts. The queries, qrels, and corpora live in |
| [`mahtaa/trqa`](https://huggingface.co/datasets/mahtaa/trqa). |
|
|
| ## Repo structure |
|
|
| | Path | Contents | Retriever type | |
| |---|---|---| |
| | `experiment_results/` | `experiment_results.zip` — retrieval + evaluation outputs | — | |
| | `indices/ecommerce/` | Indices over the **synthetic e-commerce corpus** | | |
| | `bm25/` | Lucene index | sparse | |
| | `spladepp/` | Lucene index (SPLADE++) | learned sparse | |
| | `bge/` | `bge_Flat.index` | dense (FAISS) | |
| | `e5/` | `e5_Flat.index` | dense (FAISS) | |
| | `contriever/` | `contriever_Flat.index` | dense (FAISS) | |
| | `indices/wiki/` | Indices over the **full English Wikipedia corpus** | | |
| | `bm25/` | Lucene index | sparse | |
| | `spladepp/` | Lucene index (SPLADE++) | learned sparse | |
| | `bge/` | `bge_Flat.index` — **5 parts** | dense (FAISS) | |
| | `e5/` | `e5_Flat.index` — **4 parts** | dense (FAISS) | |
| | `contriever/` | `contriever_Flat.index` — **4 parts** | dense (FAISS) | |
| | `indices/wiki_partial/` | Indices over the **partial Wikipedia corpus** | | |
| | `e5_partial/` | `e5_Flat.index` (exact / flat) | dense (FAISS) | |
| | `bm25_partial/` | Lucene index | sparse | |
| | `spladepp_partial/` | Lucene index (SPLADE++) | learned sparse | |
| | `bge_partial/` | `bge_Flat.index` (exact / flat) | dense (FAISS) | |
| | `contriever_partial/` | `contriever_Flat.index` (exact / flat) | dense (FAISS) | |
| | `qwen3_emb_4b_partial/` | `qwen3_emb_4b_Flat.index` (exact / flat) — **2 parts** | dense (FAISS) | |
|
|
| ## Reassembling split indices |
|
|
| The large dense indices (the three full Wikipedia ones and the partial |
| Wikipedia Qwen3-Embedding-4B one) are stored as 45 GB chunks. Concatenate |
| the parts in order before use: |
|
|
| ```bash |
| cat bge_Flat.index.part_* > bge_Flat.index |
| cat e5_Flat.index.part_* > e5_Flat.index |
| cat contriever_Flat.index.part_* > contriever_Flat.index |
| cat qwen3_emb_4b_Flat.index.part_* > qwen3_emb_4b_Flat.index |
| ``` |
|
|
| `cat` with a shell glob sorts the suffixes (`part_aa`, `part_ab`, …) correctly. |
| All other indices are ready to use as-is. |
|
|
