手搓「固定输出」模型合集 · Handcrafted "Fixed-Output" GGUF Models

合集总览 · Collection overview

零训练、纯脚本写死权重的「固定输出」模型:无论输入什么,都逐字节输出同一段内容。 没有算力成本、没有训练过程——权重是一张手工构造的「链表」(1 层全零 llama 架构 + embedding / output 双矩阵对齐)。

Zero-training, hand-written weights. Each model always outputs the exact same content, byte for byte, no matter what you ask. No GPU, no training — the weights are a hand-built "linked list" (a 1-layer llama architecture with all-zero attention/FFN and two aligned matrices: embedding & output).

⚡ 一行开始 · Start here

每个模型都有一个 2bit 标准版,就在它自己文件夹的根下,不用进任何子文件夹。 Each model ships a 2-bit standard build sitting at its own folder root — no sub-folder needed.

模型 · Model 2bit 标准版 · Standard build 一行运行 · One-liner
🐦 鹈鹕骑行记 · Pelican Ride 459 KB ollama run hf.co/lijie1228/a-fixed-output-model:Q2_K
⚔️ 吕布出行记 · Lü Bu Rides 435 KB ollama run hf.co/lijie1228/a-fixed-output-model:lubu-Q2_K
# 鹈鹕骑行记(骑车)· pelican on a bicycle
ollama run hf.co/lijie1228/a-fixed-output-model:Q2_K

# 吕布出行记(骑狗)· Lü Bu on a dog
ollama run hf.co/lijie1228/a-fixed-output-model:lubu-Q2_K

已收录 · Models

🐦 pelican · 鹈鹕骑行记 · Pelican Ride

逐字节输出一个交互式 HTML 页面:戴头盔的鹈鹕骑车穿过五个场景(晴日 / 黄昏 / 星夜 / 海滨 / 细雨),页内中英切换、可暂停调速、点击响铃。 Recites an interactive HTML page: a helmeted pelican cycling through five scenes, with in-page 中文/English toggle, pause & speed controls, and a bell on click.

13 个量化档(2bit 标准版 + F16 / Q8_0 / Q4_K_M / Q4_K_S / Q3_K_M / Q2_K × 固定版 fixed / 宽版 wide)——全部实测逐字节精确(含 temperature=2 极限采样)。 13 files (the 2-bit standard build + F16 / Q8_0 / Q4_K_M / Q4_K_S / Q3_K_M / Q2_K × fixed / wide) — all verified byte-for-byte (even at temperature=2).

👉 模型卡、硬件估算与双预览截图:pelican-鹈鹕/README.md

⚔️ lubu · 吕布出行记 · Lü Bu Rides

逐字节输出一个交互式 HTML 页面:戴紫金冠、持方天画戟的吕布,因为「赤兔不在」,只好骑着一条狗穿行旷野与月下;页内中英切换、可暂停调速。 Recites an interactive HTML page: Lü Bu in a golden crown with his halberd — his famous steed 赤兔 is missing, so he rides a dog across the fields and under the moon; with in-page 中文/English toggle and pause & speed controls.

同样是 13 个量化档(2bit 标准版 + F16 / Q8_0 / Q4_K_M / Q4_K_S / Q3_K_M / Q2_K × 固定版 fixed / 宽版 wide)——48/48 用例实测逐字节精确(含 temperature=2 极限采样),最小 logit 间隔 93.05(f16 为 93.76,2bit 几乎无损)。 Also 13 files (the 2-bit standard build + F16 / Q8_0 / Q4_K_M / Q4_K_S / Q3_K_M / Q2_K × fixed / wide) — 48/48 cases verified byte-for-byte, minimum logit margin 93.05 (vs 93.76 at f16).

👉 模型卡、硬件估算与双预览截图:lubu-吕布骑狗/README.md

硬件估算 · Hardware requirements

以各模型推荐的 2bit 标准版(均为宽版 Q2_K / 85 万参数)为准,全部为 4 线程纯 CPU 实测值: Based on each model's recommended 2-bit standard build (wide Q2_K / 0.85 M params), all measured on a 4-thread CPU-only machine:

项 · Item 🐦 鹈鹕 · Pelican ⚔️ 吕布 · Lü Bu
权重 · Weights 459 KB 435 KB
参数量 · Parameters 85 万 / 0.85 M 85 万 / 0.85 M
结构 · Architecture 1 层 · n_embd 256 · ffn 256 · ctx 8192 1 层 · n_embd 256 · ffn 256 · ctx 8192
词表 · Vocab 755 723
KV cache @8192 (f16) 8 MB 8 MB
实测峰值内存 @8192 ctx 37 MB 37 MB
显卡 · GPU 不需要 · not needed 不需要 · not needed
  • 显卡:不需要,-ngl 0 纯 CPU 即可(权重 < 0.5 MB,KV cache 8 MB)。 No GPU needed — pure CPU is fine (weights < 0.5 MB, KV cache 8 MB).
  • 内存:整机留 64 MB 空闲就够(37 MB 是 llama.cpp 运行时本身)。 RAM: 64 MB free is enough (the 37 MB is mostly the llama.cpp runtime itself).
  • CPU / 系统:任意 x86-64 / ARM64,Windows / macOS / Linux 均可,无 AVX 要求。 CPU / OS: any x86-64 or ARM64, Windows / macOS / Linux, no AVX required.
  • 更小的固定版 Q2_K(鹈鹕 163 KB / 吕布 144 KB,均 12 万参数)峰值内存约 31 MB,输出与宽版一致。 The even smaller fixed Q2_K (pelican 163 KB / Lü Bu 144 KB, 0.12 M params) peaks around 31 MB with identical output.

使用 · Usage

# 🐦 鹈鹕骑行记(骑车)
ollama run hf.co/lijie1228/a-fixed-output-model:Q2_K
# ⚔️ 吕布出行记(骑狗)
ollama run hf.co/lijie1228/a-fixed-output-model:lubu-Q2_K

从零复现 · Reproduce from scratch

scripts/ 是完整流水线(无需训练、无需 GPU,纯 Python + gguf 库): scripts/ contains the full pipeline (no training, no GPU — plain Python + the gguf package):

脚本 · Script 作用 · Purpose
make_fixed_gguf.py 核心构造器:把目标 HTML 编成「链表权重」,写出 GGUF / Core builder: encodes the target HTML into linked-list weights and writes a GGUF
rebuild_all.py 一键重建全部版本 / Rebuild every variant in one shot
quant_experiment.py / quant_finalize.py 量化试验与定稿 / Quantization experiments & finalization
acceptance.py / test_q32.py 验收:逐字节比对 + logit 间隔实测 / Acceptance: byte-exact diff + measured logit margins
deploy_q32.py 部署到 Ollama 与 HF / Deploy to Ollama & HF

仓库根目录的 Modelfile / Modelfile-wide 是 Ollama 本地加载用的模板。 The Modelfile / Modelfile-wide at the repo root are Ollama templates for local loading.

三个平台互挂 · Three platforms, cross-linked

同一套模型、同一批文件,三个平台互挂、内容逐字节相同(GitHub 版额外含完整流水线源码): Same models, same files, cross-linked across all three platforms — byte-for-byte identical (GitHub additionally ships the full source pipeline):

平台 · Platform 地址 · URL 特色 · What it's for
📺 哔哩哔哩 · Bilibili https://www.bilibili.com/video/BV1PmJ26xERe 42 秒宣传片「真的跑一遍给你看」· 42-second promo video
🤗 Hugging Face https://huggingface.co/lijie1228/a-fixed-output-model 模型托管,Ollama 可 hf.co 直拉 · model hosting, Ollama-pullable
🐙 GitHub https://github.com/lijie-gh/a-fixed-output-model 模型镜像 + scripts/ 全套源码 · model mirror + full scripts/ pipeline
📦 GitHub Releases pelican-v1 · lubu-v1 每个模型 13 个 GGUF 附件,英文文件名,直链下载 · 13 GGUF assets per model, direct download

整活作品,权当好玩。 · A meme project — just for fun.

Downloads last month
765
GGUF
Hardware compatibility
Log In to add your hardware

2-bit

3-bit

4-bit

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support