RedBlackBench Qwen3-14B SFT v2

Training Details

• Base: Qwen/Qwen3-14B
• LoRA: r=128, alpha=256 (improved from r=64)
• Data: 10,475 samples (deduplicated)
• LR: 5e-5 (optimized)
• Epochs: 3
• Hardware: NVIDIA H200

Improvements over v1

• LoRA rank increased: 64 → 128
• Learning rate optimized: 3e-5 → 5e-5
• Dataset cleaned: removed 6,146 duplicates
• Training data: 5,173 → 10,475 samples

Performance

• Overall reasoning: 90% (vs 80% base)
• Strategic reasoning: 100% (vs 75% base)

Usage