snowman0919/qwen38-executor-27b-checkpoints-v2

Qwen3.8-27B Executor v2

QLoRA adapter trained from the unmodified base model. Load this repository as a PEFT adapter over Qwen/Qwen3.8-27B.

• —Base revision: 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
• —Dataset revision: c4c267e27ca3cba0736f24d2c2d3997f8b12653e
• —Native maximum context: 262,144 tokens
• —Source rows: preserved unchanged; training views above 32,768 tokens use 8,192-token overlap without target-token loss
• —Quantized training: 4-bit NF4 with double quantization and BF16 compute
• —LoRA: r=32, alpha=64, dropout=0.0
• —Text: 2194 steps, 17440 deterministic rows, at least 40% Korean
• —Vision: 1275 steps over every validated vision row
• —Checkpoints: snowman0919/qwen38-executor-27b-checkpoints-v2 every 5 steps
• —Evaluation: HLE closed/web, TerminalBench 2.1, and OSWorld are run after training; OSWorld is not independent because related trajectories are intentionally included.