main_rev2_sft06

This is a Safe SFT LoRA adapter (REV2 SFT06). It uses Completion-only Training and Enhanced TOML Refinement Filter (Followup Patch).

Base Model

Qwen/Qwen3-4B-Instruct-2507

Training Data (Mixed 65:35, TOML <= 10%)

• 65%: daichira/structured-hard-sft-4k (Filtered + Refined TOML)
• 35%: u-10bei/structured_data_with_cot_dataset_512_v4 (Filtered + Refined TOML)

TOML Refinement Applied (Followup)

• Drop Audit/Log: Complete elimination of 'audit', 'timestamp', 'created by' (case-insensitive).
• Drop Repetition:
  • Quoted String Repeat (>=12 times).
  • Key Repeat ('mineral_name', etc).
  • Inline Big Array (25+ elements).

Method

• Completion-only: User prompts are masked.
• Marker: `

OUTPUT

`.

• Config: 1 Epoch, Max Seq Length 4096.