p-e-w/heretic
Training-free language model modification: Remove censorship, suppress slop, reduce sycophancy, and more
Python · ★ 39,139 · Fork 4,412 · 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
p-e-w/heretic 已在 GitHub Trending 每日榜单上榜 21 天,最高排名第 1,首次上榜 2026-02-07,最近一次上榜 2026-11-04。
| 日期 | 排名 | Star | 当日增长 |
|---|---|---|---|
| 2026-11-04 | #6 | 39,139 | +701 |
| 2026-11-03 | #6 | 38,742 | +699 |
| 2026-11-02 | #15 | 38,050 | +404 |
| 2026-09-01 | #13 | 29,862 | +537 |
| 2026-08-31 | #5 | 29,601 | +536 |
| 2026-08-30 | #5 | 29,120 | +485 |
| 2026-08-29 | #6 | 28,648 | +150 |
| 2026-06-02 | #7 | 23,215 | +249 |
| 2026-06-01 | #7 | 23,005 | +241 |
| 2026-05-28 | #7 | 22,200 | +211 |
| 2026-05-27 | #6 | 21,978 | +314 |
| 2026-03-17 | #9 | 15,496 | +788 |
| 2026-03-16 | #9 | 15,239 | +787 |
| 2026-03-15 | #7 | 14,598 | +1,066 |
| 2026-03-14 | #4 | 13,634 | +661 |
| 2026-02-20 | #8 | 8,648 | +652 |
| 2026-02-19 | #8 | 8,451 | +657 |
| 2026-02-18 | #2 | 7,943 | +947 |
| 2026-02-17 | #1 | 7,101 | +666 |
| 2026-02-08 | #4 | 4,845 | +309 |
| 2026-02-07 | #4 | 4,690 | +61 |
同期上榜的项目
- NVIDIA/auto-ontology – Auto Ontology is a Python framework for building semantic, graph-based reasoning layers over data. It enables deterministic query generation, better context retrieval, and more explainable AI for LLM and agent workflows.
- max-sixty/worktrunk – Worktrunk is a CLI for Git worktree management, designed for parallel AI agent workflows
- mattpocock/skills – Skills for Real Engineers. Straight from my .agents directory.
- floci-io/floci – Light, fluffy, and always free - The AWS Local Emulator alternative
- thedotmack/claude-mem – Persistent Context Across Sessions for Every Agent – Captures everything your agent does during sessions, compresses it with AI, and injects relevant context back into future sessions. Works with Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode + More
- vercel-labs/agent-desktop – Computer use CLI for AI agents
Heretic 项目深度剖析与商业评估报告
Heretic 项目深度剖析与商业评估报告
报告定位:面向资深工程师与架构决策者的技术尽职调查。本报告基于仓库元数据、README 公开信息及行业技术背景进行静态分析,不涉及对源码的逐行审计。
---
1. 核心定位与商业落地价值
1.1 根本痛点:安全对齐的“过拟合”与能力损耗
当前主流 LLM 的安全对齐(Safety Alignment)本质上是一种基于人类反馈的行为约束训练,其副作用是导致模型在敏感但合法的场景(如医疗咨询、历史研究、安全审计)下产生过度拒绝(Over-refusal)。传统去审查手段(如 SFT 微调、LoRA 注入)需要大量标注数据和算力,且极易破坏模型原有能力。
Heretic 直击这一痛点:通过推理时的权重正交化(Directional Ablation),在不进行任何梯度更新(Post-training)的前提下,实现“手术刀式”的定向能力移除。其核心价值主张是 “零训练成本 + 零人工调参”,将原本需要 ML 专家数周的工作压缩为一条命令行指令。
1.2 商业落地场景(非灰色地带)
在 README 中展示上榜徽章
[](https://github.fq6825.top/repo/p-e-w/heretic)