Compatibility/recovery distribution of the historical formula-review baseline. The new independent mmBERT repositories do not replace this baseline automatically. The standalone cellsentry-minicpm5-2b-cancel2-q4_k_m.gguf is 1561314592 bytes, SHA256 61d9c8ab8007dbbde0eb004089d1411fb6ec7167031f9d6fcc5b521fe8ebe49a. No separate base model or adapter is needed for this GGUF.
Base: openbmb/MiniCPM5-2B@f97400052a43d642bbc6e9975e2397e3ae6a6b52; selected formula v2 training step340, merged and Q4_K_M quantized. This is a generative formula-alert reviewer, not a general PII/extraction model or mmBERT classification artifact.
hf download OriginTone/cellsentry-minicpm5-2b-formula-v2-GGUF cellsentry-minicpm5-2b-cancel2-q4_k_m.gguf --local-dir models
The historical tested runtime is llama.cpp b11429, context1024, one finding/request, temperature0, max_tokens420, thinking disabled (enable_thinking=false). Use the embedded chat template and system message You are a spreadsheet formula auditor. Supply rule ID, address, sheet, formula, finding, rule confidence and business context. Request a one-object JSON array with cellAddress, ruleId, isValid, adjustedConfidence and reasoning.
isValid=true means the rule finding is a real error. It does not mean the formula is valid. Confidence is not calibrated. Validate JSON and identities; errors, truncation or missing evidence must not silently dismiss findings. The original full command/protocol examples remain at the fixed historical model card.
Reused external580: raw dismissal precision 65/86=75.6%, 21 real errors incorrectly dismissed. Validation-selected conservative gate: 58/75=77.3%, 17 unsafe dismissals, 12.9% coverage; the 95% external precision target was not met. Original180 regression:164/180, down from v1's170/180. Synthetic results are not independent real-workbook validation, and the gate is separate application logic. No mobile inference claim is made.
model-manifest.json, evaluation-summary.json, SHA256SUMS.txt, weight bytes, download path, LICENSE and NOTICE.txt remain unchanged. No new model benchmark was run. Model license: Apache-2.0.
这是历史公式告警复核基线的兼容/恢复分发仓;新的独立 mmBERT 不会自动替换此基线。独立文件 cellsentry-minicpm5-2b-cancel2-q4_k_m.gguf 为 1561314592 字节,SHA256 61d9c8ab8007dbbde0eb004089d1411fb6ec7167031f9d6fcc5b521fe8ebe49a,已含合并推理权重,无需再下载底座或 adapter。
底座固定为 openbmb/MiniCPM5-2B@f97400052a43d642bbc6e9975e2397e3ae6a6b52,公式v2选step340并合并为Q4_K_M。它是生成式公式告警复核,不是通用PII/抽取模型,也不是mmBERT分类工件。
下载命令见英文部分。历史验证环境为 llama.cpp b11429、context1024、每次一条发现、temperature0、max_tokens420、关闭thinking。使用内嵌模板和上述system消息,提供规则、地址、工作表、公式、发现、规则分数及业务上下文,要求含 cellAddress、ruleId、isValid、adjustedConfidence、reasoning 的单对象JSON数组。
isValid=true 表示规则发现是真实错误,不是公式有效。置信分数未经校准,必须校验JSON和身份;错误、截断或证据缺失时不得静默撤销发现。完整旧命令示例保留在固定历史模型卡。
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;验证集选择的保守门控58/75=77.3%,仍有17个不安全撤销,覆盖12.9%,未达到95%目标。原180回归164/180,低于v1的170/180。合成结果不是独立真实工作簿验证;门控属于额外应用逻辑,未声明手机推理通过。
原model-manifest、evaluation-summary、SHA256SUMS、权重字节、下载路径及LICENSE/NOTICE保持不变,本次没有新增模型效果测试。模型许可为Apache-2.0。
openbmb/MiniCPM5-2B