| 1 | model_name: "Qwen/Qwen3.5-4B" |
| 2 | # full GSM8K set is 1319 questions |
| 3 | num_questions: 1319 |
| 4 | num_fewshot: 5 |
| 5 | max_concurrency: 128 |
| 6 | max_tokens: 1024 |
| 7 | use_chat_completions: true |
| 8 | temperature: 0.7 |
| 9 | seed: 42 |
| 10 | |
| 11 | # Calibrated by running non-watermaked 13 times |
| 12 | # the threshold is the average accuracy, and the tolerance is set to 3 sigma (standard deviations) |
| 13 | accuracy_threshold: 0.885 |
| 14 | tolerance: 0.015 |
| 15 | |
| 16 | server_args: >- |
| 17 | --language-model-only |
| 18 | --max-model-len 8192 |
| 19 | --max-num-seqs 128 |
| 20 | --gpu-memory-utilization 0.4 |
| 21 | --default-chat-template-kwargs '{"enable_thinking": false}' |
| 22 | --watermark-config '{"key": 42, "context_width": 4}' |