sensix-zo/Continued-Pre-Training-Vocab-Paite

Paite Vocabulary — CPT Paragraph Text (vocab_paite_2025-12-13_paragraph.jsonl)

This file is continued pretraining (CPT) data: long plain-text sequences for causal language modeling. There is no instruction header—only a text field per line—so you can adapt token statistics and bilingual bridging patterns before instruction tuning.

Dataset composition

• Construction: Built from the same Paite vocabulary sentence pairs as the SFT release. Each underlying example uses one of three random bridging templates between quoted English and Paite spans (periods inside those quoted spans are removed in processing). Sentences are separated with ., and the stream is packed into chunks of up to 2048 cl100k_base tokens per line.
• Coverage: Broad vocabulary and short-sentence domains (daily life, food, travel, emotion, and related topics).

File description

vocab_paite_2025-12-13_paragraph.jsonl

Property	Value
Lines	284
Format	JSONL — one object per line (UTF-8)
Schema	text (string) only

Example line:

{"text": "\"The knife is very sharp\"paite pau in\"tem a hiam mahmah\"a kichi hi. ..."}

• JSON escapes: \" in the file are required so literal " appear inside the JSON string; after json.loads, the text uses normal quotes.

Technical training parameters (CPT)

• Objective: causal LM on text (standard next-token prediction).
• Learning rate: often lower than SFT (e.g. 1e-5–5e-5 — validate on loss).
• Context length: up to 2048 tokens per line by construction.
• Packing: optional; each line is already a long chunk.

Usage notes

• Format: JSONL — one JSON object per line.
• Loading: Read each line, json.loads(line)["text"], feed strings to your CPT dataloader.
• License: MIT (frontmatter); comply with your base model license (e.g. Gemma) for redistribution.

Citation

Reference this artifact by filename and date: vocab_paite_2025-12-13_paragraph.