llmcmp-distill-qwen25-7b-lora-v6q-target-no-rationale-long-ab-swap-merged

Overview

Base Model

• unsloth/Qwen2.5-7B-Instruct-bnb-4bit

Training Data

• Distilled source: safe_plus_filtered_plus_raw_partial
• Raw Kaggle train mix mode: partial
• Raw train rows are limited to non-distilled ids only
• Teacher hard mix: False
• Base-family diversity: Qwen2.5 student for ensemble comparison with Llama v5a

Training Format

• Prompt contains Prompt / Response A / Response B
• Teacher rationale is not injected into prompt-side input
• Completion is winner-only: A / B / C
• C means tie

Notes

• This model is intended for direct next-token winner inference in evaluation / submission.
• A/B/C were verified as single tokenizer tokens before training.