sensix-zo/SFT-Paite_Translation

Paite Vocabulary — SFT Translate-Only (vocab_paite_2025-12-13_translate_only.jsonl)

This dataset is a filtered slice of the Paite vocabulary instruction set for supervised fine-tuning (SFT). Only rows whose instruction starts with Translate (case-insensitive) are kept, so every example is an English-to-Paite translation task. The classic Instruction–Input–Output structure matches Alpaca-style training on Gemma, Unsloth, and similar stacks.

Dataset composition

• Task focus: English appears in input; Paite is the target in output. instruction frames the task (e.g. Translate … to Paite).
• Coverage: Broad vocabulary and short-sentence patterns (kitchen, travel, emotion, daily life, and related domains).
• Filtering: Rows whose instruction did not start with Translate were removed so supervision stays a single task type.

File description

vocab_paite_2025-12-13_translate_only.jsonl

Property	Value
Lines	26,502
Format	JSONL (one JSON object per line, UTF-8)
Schema	instruction (string), input (string), output (string)

Example line:

{"instruction": "Translate The knife is very sharp to Paite", "input": "The knife is very sharp.", "output": "tem a hiam mahmah."}

• instruction: Always begins with Translate (after optional leading whitespace).
• input: English phrase or sentence to translate.
• output: Paite translation (supervision target).

Relationship to the full release

The parent file vocab_paite_2025-12-13.jsonl may include non-translate instructions. This _translate_only file is the subset for translate-only SFT. For CPT plain-text data from the same project, see README_vocab_paite_2025-12-13_paragraph.md and vocab_paite_2025-12-13_paragraph.jsonl.

Technical training parameters (SFT)

• SFT learning rate: 2e-5 (tune if loss is unstable).
• LoRA rank (r): 64 or 128
• LoRA alpha (α): 128 or 256
• Context length: 4096 tokens is typical for Gemma/Unsloth (examples are short; packing may help throughput).
• Packing: Enable where supported (e.g. Unsloth) for faster training.

Usage notes

• Format: JSONL — one record per line.
• Structure: Each line has instruction, input, and output.
• Training: Map instruction + input to the user/prompt and train on output per your chat or Alpaca template.
• License: MIT (frontmatter); confirm compliance with your base model’s license (e.g. Gemma) before redistribution.

Citation

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