| --- |
| license: other |
| pretty_name: DR-DCI Experiment Corpora |
| tags: |
| - retrieval |
| - corpus |
| - direct-corpus-interaction |
| - browsecomp-plus |
| - wiki18 |
| - bright |
| - beir |
| --- |
| |
| # DCI experiment corpora |
|
|
| This private dataset repository stores the corpora used by the archived |
| DCI experiments, together with the scripts needed to reconstruct nested |
| scaling variants and regenerate omitted indices. |
|
|
| The repository preserves the original `corpora/...` paths. |
|
|
| ## Contents |
|
|
| | Path | Stored size | Files | Contents | |
| |---|---:|---:|---| |
| | `corpora/bcp_fineweb_nested_10m/` | 10.61 GiB | 104 | Canonical ordered 10M BrowseComp-Plus + FineWeb corpus. It contains 101 Zstandard-compressed JSONL shards, the source/composition manifest, per-shard line counts, and the build log. | |
| | `corpora/wiki18/` | 4.77 GiB | 3 | Canonical `wiki-18.jsonl.gz`, the recorded assembly sizes, and the download log. The trajectory repository contains only the requested 50-question subsets, while this is the shared retrieval corpus. | |
| | `corpora/public_ir/` | 0.55 GiB | 42 | Pinned BRIGHT Parquet snapshots and BEIR Arguana/SciFact corpora, queries, qrels, and source README files used by the relevance-ranking experiments. | |
| | `scripts/trim_nested_scaling_corpus.py` | 4 KiB | 1 | Streams the ordered 10M corpus and reconstructs the smaller nested corpus variants without requiring separately stored copies. | |
| | `scripts/rebuild_indices.md` | 2 KiB | 1 | Recorded commands and model settings for rebuilding Pyserini/Lucene, dense, and Wiki18 E5 indices. | |
|
|
| ## Nested BCP + FineWeb corpus |
|
|
| The canonical collection contains 10,000,195 documents in this order: |
|
|
| 1. 100,195 BrowseComp-Plus documents. |
| 2. 9,900,000 documents from `HuggingFaceFW/fineweb`, |
| `sample-10BT`, split `train`. |
|
|
| The stored source manifest records seed 42 and a logical shard size of |
| 100,000. The observed nested experiment boundaries are: |
|
|
| | Variant | Documents | |
| |---|---:| |
| | 100K | 100,195 | |
| | 200K | 200,197 | |
| | 400K | 400,195 | |
| | 800K | 800,195 | |
| | 10M | 10,000,195 | |
|
|
| Reconstruct a smaller variant: |
|
|
| ```bash |
| python scripts/trim_nested_scaling_corpus.py \ |
| --source corpora/bcp_fineweb_nested_10m/collection \ |
| --size 800k \ |
| --output /path/to/bcp800k |
| ``` |
|
|
| The collection shards use Zstandard level 15 with long-window mode. A |
| long-window decoder may be required: |
|
|
| ```bash |
| zstd -d --long=31 \ |
| corpora/bcp_fineweb_nested_10m/collection/docs_00000.jsonl.zst |
| ``` |
|
|
| ## Index policy |
|
|
| Generated FAISS, dense, Lucene, and Pyserini indices are intentionally not |
| stored. They can be regenerated from the canonical corpora using |
| `scripts/rebuild_indices.md`. |
|
|
| ## Integrity |
|
|
| `file_manifest.tsv` records type, size, mtime, SHA-256, relative path, and |
| symlink target for every corpus and included script. `SHA256SUMS` provides a |
| standard checksum list: |
|
|
| ```bash |
| sha256sum -c SHA256SUMS |
| ``` |
|
|
| The repository is private because the combined corpus contains material from |
| multiple upstream datasets with distinct licensing and redistribution terms. |
|
|
