qwen3-4b-h100-v5-hard-ep3

Top-ranker strategy model. Trained on H100 with a blend of three datasets (approx. 14k rows) and heavily preprocessed with custom clean_assistant_output_v2 (CoT stripping, markdown removal, TOML comment removal).

Training Configuration

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Max sequence length: 4096
• Epochs: 3
• Learning rate: 2e-5
• Effective Batch size: 32 (BS=8, GradAccum=4)
• LoRA R: 128