Historical formula-review baseline retained for compatibility, evaluation comparison and recovery. This is a generative chat LoRA/merged model, not a mmBERT scalar/classification head. It is not newly promoted or substituted by this cleanup.
openbmb/MiniCPM5-2B@f97400052a43d642bbc6e9975e2397e3ae6a6b52; do not stack the older v1 adapter.gguf/cellsentry-minicpm5-2b-cancel2-q4_k_m.gguf: 1561314592 bytes, SHA256 61d9c8ab8007dbbde0eb004089d1411fb6ec7167031f9d6fcc5b521fe8ebe49a; contains the merged inference weights.evaluation/ and hf-upload-manifest.json: existing aggregate results/provenance, unchanged.Use the saved chat template with thinking disabled, context 1024, temperature 0 and at most 420 output tokens, one alert per request. The CellSentry audit JSON field isValid=true means the finding is a real error, not that the formula is valid. Returned identities and JSON require validation; preserve original findings on failure.
Only formula review was trained in this v2; PII/extraction results were regression checks. The reused external580 benchmark recorded raw dismissal precision 65/86=75.6% (21 real errors incorrectly dismissed); a validation-selected gate recorded 58/75=77.3% (17 unsafe dismissals). It did not meet the 95% external target. New synthetic480 gate results do not establish real-world safety. This is not approval for automatic dismissal, mobile deployment, or a default-model replacement.
See the complete original model card and the unchanged evaluation directory for historical training/smoke details. No model was trained or re-evaluated for this cleanup. Weight paths and digests are preserved. Model license: Apache-2.0; preserve LICENSE and NOTICE.txt.
这是保留下载兼容、评测对照与恢复能力的历史公式复核基线,属于生成式聊天 LoRA/合并模型,不是 mmBERT 分类头;本次整理不将它重新提升为默认候选。
根目录 PEFT adapter 需配合固定底座 openbmb/MiniCPM5-2B@f97400052a43d642bbc6e9975e2397e3ae6a6b52,不能叠加旧 v1 adapter。gguf/cellsentry-minicpm5-2b-cancel2-q4_k_m.gguf 是独立合并权重,1561314592 字节,SHA256 61d9c8ab8007dbbde0eb004089d1411fb6ec7167031f9d6fcc5b521fe8ebe49a。既有 evaluation 与上传清单保持原样。
沿用保存的聊天模板,关闭 thinking,context1024、temperature0、最多420输出tokens,每次一条告警。审计 JSON 中 isValid=true 指规则发现确实是错误,不是“公式有效”;须校验身份与JSON,失败保留原发现。
v2 只训练公式复核,PII/抽取仅为回归任务。复用 external580 的原始撤销精确率 65/86=75.6%,错误撤销21个真实错误;验证集选择的门控为58/75=77.3%,仍有17个不安全撤销,未达到95%外部目标。新合成480结果不代表真实安全,也不批准自动撤销、手机部署或切换默认模型。
完整旧卡见英文部分固定链接,详细训练与smoke见原评测目录。本次没有重训或新增效果验证,权重路径和摘要不变。模型许可 Apache-2.0,须保留 LICENSE 和 NOTICE.txt。
openbmb/MiniCPM5-2B