safety-research/bloom
bloom - evaluate any behavior immediately 🌸🌱
Python · ★ 180 · Fork 89 · 在 GitHub 查看
在手机上浏览源码、订阅每日 AI 趋势日报与项目架构深度解读:获取 CopoHub 客户端(iOS & HarmonyOS)→
GitHub Trending 上榜记录
safety-research/bloom 已在 GitHub Trending 每日榜单上榜 1 天,最高排名第 7,首次上榜 2025-12-24,最近一次上榜 2025-12-24。
| 日期 | 排名 | Star | 当日增长 |
|---|---|---|---|
| 2025-12-24 | #7 | 180 | +180 |
同期上榜的项目
- rendercv/rendercv – CV/resume generator for academics and engineers, YAML to PDF
- twitter/the-algorithm – Source code for the X Recommendation Algorithm
- google/langextract – A Python library for extracting structured information from unstructured text using LLMs with precise source grounding and interactive visualization.
- vllm-project/vllm-omni – A framework for efficient model inference with omni-modality models
- stan-smith/FossFLOW – Make beautiful isometric infrastructure diagrams
- davila7/claude-code-templates – Open-source registry and CLI for Claude Code components: agents, commands, skills, hooks, MCPs, settings and mods
safety-research/bloom 项目深度剖析与商业评估报告
safety-research/bloom 项目深度剖析与商业评估报告
报告日期: 2026-08-22 | 分析对象: safety-research/bloom (v1.x, 已冻结)
---
1. 核心定位与商业落地价值
1.1 解决的根本痛点
Bloom 切入的是 LLM 评估领域一个长期被忽视的空白:动态行为评估。传统基准测试(如 MMLU、HELM)基于静态数据集,存在两个致命缺陷:一是数据污染——训练语料一旦混入测试集,指标即失真(README 中特意嵌入 Canary GUID 即为此);二是无法覆盖长尾行为——针对特定行为(谄媚、自我保存倾向、政治偏见)的测试场景需要大量人工设计,成本极高。
Bloom 的核心价值主张是:将"编写测试用例"这一过程自动化。用户只需提供行为描述(seed),系统通过 LLM 驱动的 Understanding → Ideation → Rollout → Judgment 四阶段流水线,自动生成多样化测试场景并执行评估。这使得非技术背景的安全研究员也能在数小时内完成一次行为审计,而此前这需要数周的专业工程投入。
1.2 企业级落地场景
在 README 中展示上榜徽章
[](https://github.fq6825.top/repo/safety-research/bloom)