README.md

metadata

license: odc-by
language:
  - ko
  - en
  - zh
  - ru
  - ja
pretty_name: clm_prod 5-lang monolingual BACKBONE sample (mC4)
tags:
  - clm-prod
  - byte-vocab
  - multilingual
  - backbone
  - sample

clm-backbone-5lang-sample

clm_prod 3B/7B BACKBONE corpus (step 2 DATA PREP). A bounded streaming SAMPLE — NOT the full 60B-token pretrain set (that is the step 3 H100 fire).

• Source: allenai/c4 (mC4 multilingual), configs ko/en/zh/ru/ja
• License: ODC-BY (Open Data Commons Attribution)
• real_fraction: 1.0 — every line is real downloaded web text. NO synthesis, NO LLM generation.
• Preferred set blocked: uonlp/CulturaX is GATED; the token has no data access (403 on stream) → fell back to c4.
• Format: plain newline corpus, one text line per sample, byte-vocab V=256 (UTF-8). clm_prod reads it via CLM_PROD_CORPUS.
• Balance: 4000 docs/language, round-robin interleaved (5-way balanced).

field	value
total lines	20000 (4000 x 5)
total bytes	67,733,426 (~67.7 MB)
ko bytes	16,679,968
en bytes	7,739,055
zh bytes	9,317,034
ru bytes	14,529,169
ja bytes	19,468,200
max line bytes	8192
corpus sha256	7cc86c9afd5347920fb40be48d8f22b17ca2f72e04ea172f2496b13f42df3b04

clm_prod sanity (F-CLM-PROD-DESCENT)

corpus: clm_backbone_5lang_sample.txt  (67733426 bytes, V=256)
windows: 8/8  (T=24 stride=8466675)
epoch-1 mean CE = 4.747
epoch-12 mean CE = 1.49606
F-CLM-PROD-DESCENT = 1
PASS — real-corpus mean CE descends under int4 envelope

Files: clm_backbone_5lang_sample.txt (corpus), clm_backbone_5lang.corpus.kosmos (manifest), clm_backbone_5lang.meta.json (provenance).