v0.1 — initial release with research artefacts

First release.

What's in the repo

Two end-to-end demos of "replace a neuron's role" on a frozen Gemma 4 E2B-it:

• neko/ — single-column rewrite, makes the model spontaneously say 喵
• leetcode233/ — LoRA / delta-W on a single hard problem

Everything you need to run either demo (code + small weights + dataset

• fingerprints) is in the repo itself. See README for usage.

What's attached to this release

research-artefacts-v0.1.tar.gz (133.3 MB) — three large raw tensors from
the catgirl research path that did not fit into the repo:

• step3_diff_raw.pt (45 MB) — early QA activation differential
• step4_intervene_raw.pt (186 MB) — full intervene sweep
• step4b_strong_raw.pt (217 MB) — strong-sweep variant

These are NOT required for any final demo; they are only needed if you
want to retrace the step1 → step28 catgirl research path inside
research_steps/neko_pre/. Extract with tar -xzf and move the .pt
files into research_steps/neko_pre/ so the original scripts find them.

Tarball SHA-256: FFA249E3105414E289A95E9DB186C442A2C5E236EC7D61888A248AC2A1426D7E

---

首次发布。

仓库内容

冻结 Gemma 4 E2B-it 上"替换神经元角色"的两条端到端演示:

• neko/ — 单列重写,让模型自发说"喵"
• leetcode233/ — 单道 hard 题上的 LoRA / delta-W 微调

跑两个演示所需(代码 + 小权重 + 数据集 + 指纹)全部在仓库内。用法见 README。

本 Release 的附件

research-artefacts-v0.1.tar.gz(133.3 MB)—— 猫娘研究路径里 3 个
没塞进仓库的大 raw 张量:

• step3_diff_raw.pt(45 MB)—— 早期 QA 激活差分
• step4_intervene_raw.pt(186 MB)—— 完整 intervene 扫描
• step4b_strong_raw.pt(217 MB)—— 强扫描版

最终演示都不需要它们;只有当你想顺着 research_steps/neko_pre/ 的
step1 → step28 路径重走研究流程时才用到。tar -xzf 解压后把 .pt
文件放到 research_steps/neko_pre/ 即可让原脚本找到它们。

Tarball SHA-256:FFA249E3105414E289A95E9DB186C442A2C5E236EC7D61888A248AC2A1426D7E