---
license: mit
language:
  - en
tags:
  - anima
  - corpus
  - byte-level
  - sns
  - register
size_categories:
  - 1K<n<10K
---

# anima-corpus-en-sns

anima 4-cell register corpus — **en-SNS — English casual/social conversational voice**.

Part of the anima production chat corpus = **2 languages (ko·en) × 2 registers
(general·SNS) = 4 cells** (a_chat_registers). This is one cell.

## composition (language-verified)
- **language purity:** 97.4% en (re-audited; non-target lines dropped)
- **size:** 1.33 MB · 6862 lines
- **cleaning:** per-line language classification (script + stopword), de/es/fr/ja/zh
  dropped, exact-duplicate dedup, broken/short-line drop

## language audit (provenance)
the legacy en-SNS candidate (persona_sns_corpus_5lang_v2) was only **18.7% en** by bytes (ko 21.8% · es 18.8% · fr 16.2% · de 13.9%) — a 5lang mixture. Split to en-only lines + deduped (~115K duplicate lines removed across the SNS split) → **97.4% en** (2.3% residual = short captions/hashtags, not other languages).

Method: script/charset + stopword-fingerprint per-line classifier (torch-free,
deterministic). Audit tool + reproduce script: `state/clm303_clean_corpus/`
(`langaudit.py` measures composition, `build_corpus.py` rebuilds this cell).
*(Measured by the anima clm303-retrain build, 2026-06-24; cross-referenced by the
governance rule `a_chat_registers` as the precedent that mandates per-cell language
verification.)*

## sources (repo · license)
| source | license |
|--------|---------|
| anima persona_sns_corpus_5lang_v2 (en lines only) | MIT |

## sha256
```
49f347c72416aad24d9b16dd9b406173c6d8bb764026381e52f6e36bc5b05f70  anima-corpus-en-sns.txt
```

## role in the 4-cell register
| cell | language | register |
|------|----------|----------|
| anima-corpus-ko-general | ko | general |
| anima-corpus-en-general | en | general |
| anima-corpus-ko-sns | ko | SNS |
| anima-corpus-en-sns | en | SNS |

⚠️ **KNOWN-SMALL BASELINE** — clean en-SNS is only 1.33MB (no permissive large en-SNS source exists on HF: daily_dialog/empathetic_dialogues are non-commercial, REDDIT_comments is unlicensed). Use size-proportional sampling so this cell is NOT memorized. youtube/insta-en augmentation is a tracked follow-up (a_chat_registers SNS register completion).

## why this rebuild
The prior "en" cells were silently 5lang mixtures (only ~20% en — de/es/fr/ko
contamination) and a single 4MB cell was the entire effective training corpus
(~120× repetition = memorization). These clean cells fix both: every cell is
language-verified, and training uses **size-proportional sampling** so per-cell
repetition is uniform (no small cell is over-repeated).