main_rev2_sft04

This is a Safe SFT LoRA adapter (REV2 SFT04). It uses Completion-only Training and Hard SFT Filtering.

Base Model

Qwen/Qwen3-4B-Instruct-2507

Training Data (Mixed 65:35)

• 65%: daichira/structured-hard-sft-4k (Filtered High Quality)
• 35%: u-10bei/structured_data_with_cot_dataset_512_v4 (Filtered Output-only)

Hard Filter Applied

• Length Limit (Format-wise)
• Anti-Log/Audit Keywords
• Repetition Check (Tokens, Lines, Gases)

Method

• Completion-only: User prompts are masked (-100 output label).
• Marker: `

OUTPUT

` inserted before assistant output.

• Config: 1 Epoch, Max Seq Length 4096.