Vietnamese National Legal Database (vbpl.vn) corpus with structure + embedding layers

README.md

@@ -29,14 +29,6 @@ configs:
   data_files:
   - split: train
     path: documents-*.parquet
-- config_name: embed
-  data_files:
-  - split: train
-    path: embed-*.parquet
-- config_name: reduce
-  data_files:
-  - split: train
-    path: reduce-*.parquet
 ---
 
 # Vietnamese National Legal Database — `vbpl.vn`
@@ -371,91 +363,34 @@ for sec in structure.get("sections", []):
     print(sec["kind"], sec["label"])
 ```
 
-## Companion stages · `embed` + `reduce`
-
-Alongside the default `documents-*.parquet` shards (one row per
-document, with text + structure), the repo also carries the
-**embed** and **reduce** pipeline outputs as separate parquet
-bundles. Both join back to the `documents` table on the
-`doc_name` primary key. Only the **embeddable** rows
-(147,317 after dropping NULL-markdown
-docs) appear in these stages.
-
-### `embed-*.parquet` — dense vectors
-
-15 shards (~93 MB each, ~1.33 GB total, 10 000 rows per shard,
-deterministic `doc_name` ordering). Schema mirrors the
-`anle.toaan.gov.vn` corpus's embed stage exactly so
-cross-corpus joins are straightforward:
-
-| Field | Type | Description |
-|---|---|---|
-| `doc_name` | string | Join key back to `documents-*.parquet`. |
-| `text_hash` | string | SHA-256 of the post-normalisation `markdown` (stable across re-runs). |
-| `embedding` | list<float64> | **2048-D** dense vector from `nvidia/llama-nemotron-embed-1b-v2` (default; other models give other dims). |
-| `embedding_dim` | int64 | Length of `embedding` (denormalised for fast filtering, always `2048` in this release). |
-| `embedding_model_id` | string | Model slug as the embedder backend reports it. |
-| `embedding_text_hash` | string | SHA-256 of the exact text fed to the embedder (differs from `text_hash` when sliding-window chunking applies). |
-| `embedding_chunks_used` | int64 | Number of windows mean-pooled into the final vector (1 when the doc fits in one window). |
-| `embedding_chunking` | string | Chunking strategy: `off` / `sliding` / `sentence`. |
-
-### `reduce-*.parquet` — 2-D projections + cluster ids
-
-15 shards (~0.5 MB each, ~7 MB total). PCA + t-SNE + UMAP run
-with `cfg.reducer.n_components=2` so the `*_z` columns that
-existed in the on-disk per-doc shards are dropped here.
-HDBSCAN cluster ids land in `[-1, N]` (`-1` is the noise
-bucket).
-
-| Field | Type | Description |
-|---|---|---|
-| `doc_name` / `text_hash` | string | Join keys back to `documents-*.parquet` and `embed-*.parquet`. |
-| `pca_x` / `pca_y` | float64 | 2-D PCA projection of the 2048-D embedding. |
-| `umap_x` / `umap_y` | float64 | 2-D UMAP projection (the one used in the scatter PNGs above). Fit globally across all rows for cross-doc comparability. |
-| `cluster_id` | int64 | HDBSCAN cluster label; `-1` is the noise / unclustered bucket. |
-
-Quick load (each stage is a `data_files` glob; the default
-`load_dataset("tmquan/vbpl-vn")` still resolves to the
-`documents` config):
-
-```python
-from datasets import load_dataset
-
-embed = load_dataset(
-    "tmquan/vbpl-vn",
-    data_files="embed-*.parquet",
-    split="train",
-)
-print(embed[0]["doc_name"], len(embed[0]["embedding"]))
-# e.g. "100000 2048"
-
-reduce = load_dataset(
-    "tmquan/vbpl-vn",
-    data_files="reduce-*.parquet",
-    split="train",
-)
-print(reduce[0]["doc_name"], reduce[0]["umap_x"], reduce[0]["cluster_id"])
-# e.g. "100000 1.7142 -1"
+## Embedding + reduction artefacts
+
+The Hub ships the **`documents` config only**
+(`documents-*.parquet`, one row per document, with text +
+structure). The 2048-D
+`nvidia/llama-nemotron-embed-1b-v2` embeddings, the global UMAP / PCA
+coordinates, and the HDBSCAN cluster ids are computed during the
+build (over 147,317 embeddable rows,
+i.e. all documents with non-null `markdown`) and used to render
+the scatter PNGs below, but they are **not bundled as separate
+parquet configs on the Hub** -- 1.3 GB of dense vectors per
+re-build is too costly to ship for a corpus this size when the
+embeddings are deterministic from `markdown` + a model id.
+Re-derive the same matrices locally via:
+
+```bash
+git clone https://github.com/<owner>/ViLA   # the build repo
+python -m packages.datasites.vbpl --pipeline embed
+python -m packages.datasites.vbpl.\_reduce\_inproc   # global UMAP
 ```
 
-To join embed-stage vectors back to the document metadata, do
-the join client-side on `doc_name`:
-
-```python
-import pandas as pd
-
-docs   = load_dataset("tmquan/vbpl-vn", split="train").to_pandas()
-embed  = load_dataset("tmquan/vbpl-vn",
-                      data_files="embed-*.parquet",
-                      split="train").to_pandas()
-joined = docs.merge(embed, on="doc_name", how="inner")
-# joined now has the title / so_hieu / markdown / ... columns
-# next to the 2048-D embedding vector for every embeddable doc.
-```
+which yields `parquet/embeddings/<doc_name>.parquet` and
+`parquet/reduced/<doc_name>.parquet` per-doc shards keyed back to
+`documents.doc_name`.
 
 ## Trực quan hoá embedding · Embedding visualization
 
-Mỗi điểm là một văn bản pháp luật; toạ độ là vector embedding 2048-D từ `nvidia/llama-nemotron-embed-1b-v2` chiếu xuống 2D bằng UMAP, cụm bằng HDBSCAN. Sáu mặt phân hoạch: `scope`, `doc_type` (mã ngắn), `legal_type` (tên đầy đủ), `legal_area` (lĩnh vực pháp luật), `year`, `cluster_id`. Các nhãn tail-end (sau top-18) được dồn vào nhóm *Khác / Other* màu xám để chú giải đọc được. — Each dot is one legal document; coordinates are the 2D UMAP projection of a 2048-D embedding from `nvidia/llama-nemotron-embed-1b-v2`, with HDBSCAN cluster ids. Six facets: `scope`, `doc_type` (canonical short code), `legal_type` (canonical full name), `legal_area` (subject domain), `year`, `cluster_id`. Tail-end labels beyond the top 18 are collapsed into a grey *Khác / Other* bucket to keep the legend legible.
+Mỗi điểm là một văn bản pháp luật; toạ độ là vector embedding 2048-D từ `nvidia/llama-nemotron-embed-1b-v2` chiếu xuống 2D bằng **UMAP fit globally** trên toàn bộ kho văn bản (không chia batch) để toạ độ so sánh được giữa các tài liệu. Năm mặt phân hoạch: `scope`, `doc_type` (mã ngắn), `legal_type` (tên đầy đủ), `legal_area` (lĩnh vực pháp luật), `year`. Các nhãn tail-end (sau top-18) được dồn vào nhóm *Khác / Other* màu xám để chú giải đọc được. — Each dot is one legal document; coordinates are the 2D UMAP projection of a 2048-D embedding from `nvidia/llama-nemotron-embed-1b-v2`, **fit globally across the whole corpus** so positions are directly comparable across documents. Five facets: `scope`, `doc_type` (canonical short code), `legal_type` (canonical full name), `legal_area` (subject domain), `year`. Tail-end labels beyond the top 18 are collapsed into a grey *Khác / Other* bucket to keep the legend legible.
 
 ### UMAP colored by `scope`
 
@@ -503,9 +438,17 @@ intercepts the resulting authenticated XHRs, downloads any
 4. **Embed** -- `nvidia/llama-nemotron-embed-1b-v2` (2048-D) over the
    normalised markdown; sliding-window mean pool when a doc exceeds
    the model's native context window.
-5. **Reduce** -- PCA + t-SNE + UMAP on the embedding matrix +
-   HDBSCAN cluster ids. cuML on a GPU worker; sklearn / umap-learn
-   / hdbscan otherwise.
+5. **Reduce** -- PCA + UMAP on the embedding matrix + HDBSCAN
+   cluster ids. The May-2026 rerun switched from per-batch
+   reduction (each Curator `DocumentBatch` independently) to a
+   **global in-process fit** over the full 147,317
+   × 2048-D matrix so the projection is comparable
+   across documents (`packages/datasites/vbpl/_reduce_inproc.py`).
+   t-SNE was dropped from the rerun because a single global fit
+   under `random_state=0` (single-threaded by construction) took
+   prohibitively long on this corpus; PCA + UMAP cover the same
+   visual story without it. cuML on a GPU worker; sklearn /
+   umap-learn / hdbscan otherwise.
 
 All five layers are deterministic and re-runnable; re-running any
 stage with the same `--limit` is a no-op (each stage skips
@@ -529,7 +472,7 @@ embedding parquets for the NULL-markdown rows are dropped from
 the embedding corpus entirely so the reducer fits on the
 embeddable rows only (147,317-row corpus).
 
-Captured: `2026-05-23T14:10:30.764933+00:00`.
+Captured: `2026-05-23T14:29:39.492383+00:00`.
 
 ## Nguồn · Source
 

manifest.json

@@ -536,5 +536,5 @@
       "share": 0.07243958645527698
     }
   },
-  "completed_at": "2026-05-23T14:10:30.764933+00:00"
+  "completed_at": "2026-05-23T14:29:39.492383+00:00"
 }