build_identity dict | completed_rows int64 | created_at string | dimension int64 | embedding_complete bool | resolved_model_revision string | total_rows int64 | truncation_estimate dict | updated_at string |
|---|---|---|---|---|---|---|---|---|
{
"corpus": {
"require_id": true,
"row_mapping": "zero-based JSONL line; native id stored in source",
"sha256": "c3f83ca66a42c2dc54455dc0b32ea679b7f9a1738434fe31164cdebb33265d03",
"size_bytes": 36855391027,
"tar_member": null,
"text_format": "contents field exactly as stored in TRQA JSONL; v2"... | 57,745,780 | 2026-11-04T11:14:19.668701+00:00 | 768 | true | 9bbca17d9273fd0d03d5725c7a4b0f6b45142062 | 57,745,780 | {
"estimated_truncation_rate": 0.020470642301038063,
"max_observed_tokens": 23035,
"sampled_passages": 1183744,
"would_truncate": 24232
} | 2026-11-05T11:06:38.176466+00:00 |
{
"corpus": {
"require_id": true,
"row_mapping": "zero-based JSONL line; native id stored in source",
"sha256": "c3f83ca66a42c2dc54455dc0b32ea679b7f9a1738434fe31164cdebb33265d03",
"size_bytes": 36855391027,
"tar_member": null,
"text_format": "contents field exactly as stored in TRQA JSONL; v2"... | null | 2026-11-04T11:14:19.668701+00:00 | null | null | 9bbca17d9273fd0d03d5725c7a4b0f6b45142062 | null | {
"estimated_truncation_rate": 0.020470642301038063,
"max_observed_tokens": 23035,
"sampled_passages": 1183744,
"would_truncate": 24232
} | null |
TRQA Wikipedia GTE embeddings
Build in progress. Embeddings are written in source order under wiki-gte/embedding-shards/; each filename gives its zero-based half-open row interval. Shards are float32 NumPy arrays of shape (rows, 768). Each embedding corresponds to the line at the same zero-based position in the original source corpus.
Source: mahtaa/trqa, data/shared/corpus_wikipedia_en-20251001_rewritten.jsonl, revision 71e8c916222958c964c74321f31de978af76f4e2, SHA-256 c3f83ca66a42c2dc54455dc0b32ea679b7f9a1738434fe31164cdebb33265d03. The corpus contains 57,745,780 passages.
Model: Alibaba-NLP/gte-multilingual-base, revision 9bbca17d9273fd0d03d5725c7a4b0f6b45142062; SentenceTransformer encoding, fp16 model, 256-token limit, L2-normalized embeddings saved as float32. The input for each embedding is exactly the value of the contents field in the source record, without a title or any other text added; model tokenization and truncation still apply. These are normalized GTE corpus vectors for the FullWiki direct-search pipeline; they are not vectors of a trained Q-RAG action tower.
Native TRQA passage IDs remain in the source corpus and are not zero-based row numbers. wiki-gte/build-state.json is uploaded only after the whole corpus is covered; until then the dataset is incomplete.
- Downloads last month
- 801