Upload README.md with huggingface_hub

README.md

@@ -36,55 +36,97 @@ tags:
   - corpus
   - cleaned
   - pretraining
-pretty_name: Scraped-Data (Cleaned Wikipedia Corpus)
+  - internet-archive
+  - gutenberg
+pretty_name: Scraped-Data — 1TB in Progress
 ---
 
-# Scraped-Data
+# Scraped-Data — Building to 1 TB
 
-A multi-language corpus of **~8.8 million** cleaned Wikipedia article records,
-scraped, parsed, and cleaned at scale on a CPU-only Google Colab session.
+A continuously-growing, cleaned text corpus scraped, parsed, cleaned,
+**losslessly-zipped**, and auto-pushed from a CPU-only Google Colab pipeline.
 
-## Files
+**Current repo size: ~26 GB → target: 1 TB** (incremental batch uploads).
+
+## Current files
 
 | File | Size | Contents |
 |---|---|---|
-| `corpus_batch1.clean.zip` | 4.80 GB | ~2.9M docs |
+| `corpus_batch1.clean.zip` | 4.80 GB | ~2.9M Wikipedia docs (stored, from first scrape) |
 | `corpus_batch2.clean.zip` | 4.85 GB | ~2.5M docs |
 | `corpus_batch3.clean.zip` | 4.91 GB | ~2.3M docs |
+| `corpus_batch1.v3.zip` | 4.37 GB | de-wiki shards, deflate-level-6 |
+| `corpus_batch2.v3.zip` | 3.74 GB | fr + ru shards |
+| `corpus_batch3.v3.zip` | 3.45 GB | es + ru shards |
+| `manifest.jsonl` | — | durable dedup ledger (zip → bytes → source keys) |
+
+Next batches are pushed live as they finish. Total so far: **26.1 GB** compressed,
+~8.8M+ records (+ additional Gutenberg/IA batches incoming).
+
+> **Live status:** batch 4 scraping (2.5 min in, ~3.1 GB raw, ~748k docs).
+> The Colab scraper runs continuously: ~10.6 GiB raw per batch → losslessly zipped
+> (ZIP_DEFLATED, level 6, allowZip64), uploaded immediately via `huggingface_hub`,
+> with a manifest for cross-VM resume.
+
+## Sources
 
-Each archive contains a single JSONL file (`clean.jsonl`), one JSON object per line.
+- **Wikipedia** — official `wikimedia/wikipedia` `20231101` parquet shards (HF mirror
+  of Wikimedia dumps); mirror `mirror.accum.se` as fallback. Multi-language
+  priority queue: de, fr, ru, es, it, ja, zh, pt, ar, fa, tr, pl, nl, id, ko,
+  vi … (300+ configs; enwiki is only one shard of the total).
+- **Internet Archive** — `collection:gutenberg` (public-domain plain-text books)
+  via `advancedsearch` + `metadata` + `archive.org/download` — polite
+  `IA_CONCURRENCY=20`, per-host pacing, robots-friendly.
 
-## Schema
+Both pipelines emit the same JSONL record format and feed the same batch/zip
+loop.
+
+## Record schema (JSONL, one object per line)
 
 ```json
-{"src": "wiki", "url": "...", "title": "...", "chars": 1234, "text": "..."}
+{"src": "wiki", "url": "https://...", "title": "...", "chars": 1234, "text": "..."}
+{"src": "ia",   "url": "https://archive.org/details/...", "title": "...", "chars": 1234, "text": "..."}
 ```
 
-- `src` — source type (`wiki`)
-- `url` — canonical article URL
-- `title` — article title (trimmed)
+- `src` — `wiki` or `ia`
+- `url` — canonical article/item URL
+- `title` — title trimmed
 - `chars` — character count of `text`
 - `text` — cleaned plain-text body
 
-## Provenance
-
-Sourced from the official `wikimedia/wikipedia` dump snapshot `20231101`
-(parquet shards), which itself derives from Wikipedia dumps published by the
-Wikimedia Foundation. Underlying articles are © their respective Wikipedia
-contributors and are dual-licensed under CC-BY-SA; this compilation is
-additionally distributed under ODATL-1.0 (below). Attribution for upstream
-content must be preserved by downstream users in accordance with both
-licenses.
-
 ## Cleaning pipeline
 
-1. Segmented parallel HTTP download of source shards (~200 Mbps sustained)
-2. Multi-process zstd/parquet decode with orjson serialization
-3. Per-record cleaning: control-character stripping, tab/NBSP removal,
-   CRLF normalization, blank-line collapse, email redaction (`[email]`)
-4. Minimum length gate (280 chars), empty-text drop
-5. Title+URL exact-duplicate removal across each batch
-6. Validation pass: every emitted record is valid UTF-8 JSON
+1. Segmented parallel HTTP of source shards (`SEGMENTS=10`, `SEGMENTS=10` for
+   parquet, 8 concurrent `archive.org` fetches)
+2. Multi-process (2× fork) decode: `pyarrow`/`orjson` for parquet, streaming
+   `bz2` for legacy dumps, plain-text for IA
+3. Per-record `clean_wiki` / IA Gutenberg-header strip:
+   - Gutenberg header/footer carve (`*START…*END*`), tab/NBSP→space,
+     CRLF normalization, control-char strip, blank-run collapse,
+     email redaction (`[email]`)
+   - Min-length gates: `MIN_TEXT_LEN=300`, `MIN_DOC=280`
+   - Exact-duplicate removal per batch (title+URL `blake2b-12`)
+4. Batch-shard JSONL → ZIP (deflated level 6, `allowZip64`), immediate
+   `HfApi.upload_file` + manifest append + manifest re-upload for resume
+
+All ZIPs are **lossless** (`ZIP_DEFLATED` is LZ77+Huffman — perfectly
+reversible).
+
+## Provenance & attribution
+
+- Underlying Wikipedia articles © their contributors, dual-licensed
+  CC-BY-SA/GFDL.
+- Underlying IA Gutenberg texts are public domain (U.S.).
+- This compilation and its cleaned artifacts are additionally distributed
+  under **ODATL-1.0** (below). Users must preserve *both* upstream and
+  compilation attribution.
+
+## Reproducibility
+
+The full scraper (`scraper_v3.py`) runs entirely on Google Colab
+(CPU-only). State is kept in a `manifest.jsonl` on the HF repo itself;
+a new VM resumes by skipping `done_keys` (shard URLs + `ia:<id>`).
+Progress and logs live in `/content/data/progress.json` & `scraper.log`.
 
 ## License