Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,76 @@
+---
+license: odc-by
+language:
+- en
+- fr
+- de
+- es
+- ko
+pretty_name: anima 5-lang 7B web-scale corpus (R2-staged pointer index)
+tags:
+- byte-level
+- web-scale
+- fineweb
+- multilingual
+- r2-staged
+size_categories:
+- 100B<n<1T
+---
+
+# anima-corpus-5lang-7b-webscale
+
+> **⚠️ R2-STAGED, NOT HF-HOSTED.** The **143.60 GiB** of raw byte-corpus lives **only in
+> Cloudflare R2** (bucket `phanes`, prefix `anima-7b/web/`) — an egress-free durable staging
+> store. **This HF dataset holds the manifest + card + 200 KB per-language sample heads, NOT
+> the 143 GB.** The full corpus is **fetched from R2** for training; it is not HF-mirrored and
+> not held locally.
+
+The **web-scale 7B-Chinchilla-optimal** extension of the anima default-lane corpus. Byte-level
+(V = 256), 5-language balanced (en / fr / de / es / ko), web-derived under **ODC-BY**.
+
+| metric | value |
+|---|---|
+| total byte-corpus | **143.60 GiB** (154.19 GB decimal) |
+| tokens / param (7B) | **22.0 tok/param** (Chinchilla-optimal 20 → **EXCEEDS**) |
+| fraction of 7B-optimal | **110.1%** of the 140B-token budget |
+| byte vocabulary | **V = 256** |
+| shards | 20 |
+| license | **ODC-BY** (web-derived) |
+
+## Per-language / per-source
+
+| lang | GB (decimal) | source (ODC-BY) |
+|---|---:|---|
+| en | 25.83 | `HuggingFaceFW/fineweb` sample/10BT |
+| fr | 32.79 | `HuggingFaceFW/fineweb-2` fra_Latn |
+| de | 31.00 | `HuggingFaceFW/fineweb-2` deu_Latn |
+| es | 32.94 | `HuggingFaceFW/fineweb-2` spa_Latn |
+| ko | 31.64 | `HuggingFaceFW/fineweb-2` kor_Hang |
+| **total** | **154.19** | — |
+
+**Korean web bulk is INCLUDED** (≈ 31.64 GB) — this fixes the prior clean-license build's
+Korean gap (Project Gutenberg has no Korean, so the earlier `gb-balanced` corpus had ko art /
+consciousness = 0 and leaned wiki-only).
+
+## License & hygiene
+
+- **ODC-BY**, Common-Crawl-derived FineWeb / FineWeb-2. Attribution per ODC-BY.
+- **PII-scrubbed**: `email → [EMAIL]`, `phone → [PHONE]` (these tokens are the scrub markers
+  and are expected in the text). **Control bytes 0xFE / 0xFF stripped** (V = 256).
+- **No persona / role injection.** anima's persona + identity voice stay **authored**
+  elsewhere; this is the web factual/linguistic base only.
+
+## How to fetch from R2
+
+`MANIFEST.json` carries the per-shard R2 keys + bytes + sha256. Stream each shard from R2
+(`https://<account>.r2.cloudflarestorage.com`, region `auto`, bucket `phanes`) with boto3 and
+verify against the per-shard sha256. See `CORPUS_CARD.md` for the full code snippet.
+
+## Scope (honest)
+
+This is the **CORPUS** — an R2-staged pointer index. The actual **7B train is a separate
+follow-on GPU fire**; **no 7B-trained-model is claimed** here. Numbers (143.60 GiB / 22.0
+tok/param / 110.1% / per-lang) are the achieved corpus values verified against the R2 manifest
+byte counts.
+
+See `CORPUS_CARD.md` for the full card.