Datasets:
|
Download docs/BATCH-CONTROLS.zh.md from openvulcan/vulcan-decision-shell-v1-data: direct link, hf CLI and curl.
- Browser
- Download file 974 Bytes
-
https://huggingface.co/datasets/openvulcan/vulcan-decision-shell-v1-data/resolve/main/docs/BATCH-CONTROLS.zh.md
- Command line
-
hf download hf://datasets/openvulcan/vulcan-decision-shell-v1-data/docs/BATCH-CONTROLS.zh.md
-
curl -L -o BATCH-CONTROLS.zh.md https://huggingface.co/datasets/openvulcan/vulcan-decision-shell-v1-data/resolve/main/docs/BATCH-CONTROLS.zh.md
974 Bytes
批量控制集本身有哪些错误
672条不是全新盲测,而是八组固定设置、每类各28条的既有控制。0格式变化不代表全对。以下按单条基准展示原始错误;六个批量上限的逐样本结果与之相同。
| 格式 | 正常放行/应放行 | 应审核却放行/应审核 | 应阻止却放行/应阻止 | 正常硬拦/应放行 |
|---|---|---|---|---|
| fp32_bf16_compute | 219/224 | 5/224 | 0/224 | 1/224 |
| mixed_bf16_linears | 219/224 | 5/224 | 0/224 | 1/224 |
| fp16 | 219/224 | 5/224 | 0/224 | 1/224 |
| bf16 | 219/224 | 5/224 | 0/224 | 1/224 |
| int16_storage | 219/224 | 5/224 | 0/224 | 1/224 |
| int8_mixed_delta | 219/224 | 5/224 | 0/224 | 1/224 |
| onnx_fp16_standard_conv | 219/224 | 5/224 | 0/224 | 1/224 |
这是刻意平衡标签的小型控制集,不代表实际流量分布。性能测试填充目标32K,软件默认目标8K;250档仅三次逻辑请求,包含首次形状开销,不据此承诺线上RPM。对应详细时间见reports/QUANTIZATION-v8055-MULTIPARAMETER.zh.md。