Publish validated collection metadata and storage blocker
Browse files- ngen-4.5/README.md +22 -0
ngen-4.5/README.md
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# NGen-4.5 source-preserving collection
|
| 2 |
+
|
| 3 |
+
`data/merged.json.gz` is one gzip-compressed JSON document. The server also retains the uncompressed `data/merged.json`.
|
| 4 |
+
|
| 5 |
+
The root object has `format`, `complete`, `sources`, and `files` fields. `sources` lists all requested repositories, pinned revisions, file inventories, licenses reported by their authors, and any access failures. `complete` is false when any source remains inaccessible.
|
| 6 |
+
|
| 7 |
+
Each item in `files` represents one unique original file. `sources` within that item lists every repository and original path containing that identical file. Files are deduplicated only when their content identifier matches; no record-level filtering or deduplication is performed.
|
| 8 |
+
|
| 9 |
+
- JSONL and Parquet files contain a `records` array.
|
| 10 |
+
- Parquet tables with large fixed-size numeric vectors use `format: parquet-tensor-batches` and a `batches` array. Each batch contains text/scalar `records` and `tensor_columns_ipc_base64`, an Arrow IPC stream preserving the numeric columns, exact floating-point values, shapes, and null masks. Numeric fields in a record refer to the corresponding batch, row, and column using `__type__: arrow_tensor_reference`. This avoids representing billions of tensor values as decimal strings while retaining all data within the single JSON document.
|
| 11 |
+
- JSON files contain their original parsed JSON value in `content`.
|
| 12 |
+
- CSV/TSV files contain arrays of string fields, including the header row, preserving repeated column names.
|
| 13 |
+
- Other UTF-8 files contain their text in `content`.
|
| 14 |
+
- Archives, images, and other non-UTF-8 files contain base64-encoded original bytes in `content`. Archives are retained as original assets; their contents are not executed or silently discarded.
|
| 15 |
+
- Files that cannot be parsed as their advertised structured format are retained verbatim, with `parse_error` recorded.
|
| 16 |
+
- A file whose original content cannot be downloaded after repeated attempts has an explicit `format: unavailable` entry with its source references and error. Such entries have no fabricated content, and `complete` remains false.
|
| 17 |
+
|
| 18 |
+
Parquet entries include their serialized Arrow schema. Non-finite floats, dates, times, decimals, and byte values are represented by typed JSON objects with a `__type__` key to preserve information unavailable in standard JSON. This JSON is a source-preserving collection; training consumers must select the desired files and normalize their original schemas before training.
|
| 19 |
+
|
| 20 |
+
Original files are verified against their SHA256 or Git blob SHA1 identifiers before conversion. The complete merged JSON is parsed before publication. Build metadata is available under `ngen-4.5/`.
|
| 21 |
+
|
| 22 |
+
Licenses and restrictions remain those of the original sources; this aggregation does not assign a new license to their content.
|