webmp3/Sakura-SmolLM2-135M-Instruct-GGUF

Sakura · GGUF quantization of HuggingFaceTB/SmolLM2-135M-Instruct

Independent community quantization. Not an official release of the upstream author.

1. Model Overview

Architecture: llama · source formats: gguf.

Original: 0.3 GB (BF16 GGUF, the source of these quantizations).

Quant	Size	Compared with the original
Q4_K_M	0.1 GB	61 % smaller

2. Original Model / Attribution

• Original model: HuggingFaceTB/SmolLM2-135M-Instruct, commit 671ad66db42a400967d6efc49ea74888ccf68936 (model card at that commit)
• Submitted repository: webmp3/Sakura-SmolLM2-135M-Instruct-BF16-GGUF
• Declared base models: HuggingFaceTB/SmolLM2-135M-Instruct (relation: quantized)
• Quantized from: webmp3/Sakura-SmolLM2-135M-Instruct-BF16-GGUF at 671ad66db42a400967d6efc49ea74888ccf68936, file Sakura-SmolLM2-135M-Instruct-BF16.gguf (SHA-256 e8dbe545492a548832dbfd8fc178fa61724aaadd9916115da636871a84f1ee46; provenance verified: True)
• Quantized by webmp3. Upstream authors retain their license and attribution.

3. Files

File	Bytes	SHA-256	Quant
Sakura-SmolLM2-135M-Instruct-BF16-Q4_K_M.gguf	105454656	245b1a6d33d63cb6851417b9945390e084059d01a8a221bae4352d5fc5d2b49e	Q4_K_M

4. How These Were Made

This route performs existing-GGUF requantization from the pinned BF16 source GGUF named above. This job did not convert the upstream BF16/safetensors checkpoint.

No importance-matrix calibration was run by this job.

The quant label names the llama.cpp preset. Stock llama.cpp keeps some tensors (for example normalization weights) in higher precision, so not every tensor is stored in the named type; the table shows what each file really contains.

File	Tensors by type
Sakura-SmolLM2-135M-Instruct-BF16-Q4_K_M.gguf	F32: 61, Q4_K: 16, Q5_0: 166, Q6_K: 14, Q8_0: 15

5. Verification

• GGUF structure valid: True
• Real inference test (every file loaded by llama.cpp and used to generate text once): yes

File	llama.cpp build	Executable SHA-256
Sakura-SmolLM2-135M-Instruct-BF16-Q4_K_M.gguf	b11669-620d2225c	3b7167843ffb0f3677be964dc80cfc59a0167c82e11eff62f6334f66e831f209

6. Download and Usage

llama-cli -m <file>.gguf -p "Hello" -n 32

Needs a llama.cpp build that supports this architecture. If a quantization is split into numbered parts, keep them together and load the first one.

7. Limitations

No coding benchmark, perplexity or quality score is claimed unless a measured value is listed above.

8. License Information

Source license: apache-2.0. apache-2.0 allows redistribution with attribution

中文说明 · 樱花 Sakura（点击展开）

1. 概述

独立的社区量化版本，并非上游作者的官方发布。

架构：llama · 源格式：gguf。

原始模型：0.3 GB（BF16 GGUF，这些量化版本的来源）。

量化	大小	相比原始模型
Q4_K_M	0.1 GB	小 61 %

2. 原始模型与署名

• 原始模型：HuggingFaceTB/SmolLM2-135M-Instruct，提交 671ad66db42a400967d6efc49ea74888ccf68936（该提交的模型卡）
• 声明的基础模型：HuggingFaceTB/SmolLM2-135M-Instruct（关系：quantized）
• 量化来源：webmp3/Sakura-SmolLM2-135M-Instruct-BF16-GGUF，提交 671ad66db42a400967d6efc49ea74888ccf68936，文件 Sakura-SmolLM2-135M-Instruct-BF16.gguf（SHA-256 e8dbe545492a548832dbfd8fc178fa61724aaadd9916115da636871a84f1ee46；出处已验证：True）
• 量化者：webmp3。上游作者保留其许可证与署名。

3. 文件

文件	字节数	SHA-256	量化
Sakura-SmolLM2-135M-Instruct-BF16-Q4_K_M.gguf	105454656	245b1a6d33d63cb6851417b9945390e084059d01a8a221bae4352d5fc5d2b49e	Q4_K_M

4. 制作方式

该路线从上方标明的固定版本 BF16 源 GGUF 进行现有 GGUF 的再量化。本任务没有转换上游的 BF16/safetensors 检查点。

本任务没有运行重要性矩阵（imatrix）校准。

量化标签指 llama.cpp 的预设名称。标准 llama.cpp 会把部分张量（例如归一化权重）保留为更高精度，因此并非每个张量都以所标注的类型存储；下表列出每个文件实际包含的内容。

文件	按类型的张量数
Sakura-SmolLM2-135M-Instruct-BF16-Q4_K_M.gguf	F32: 61, Q4_K: 16, Q5_0: 166, Q6_K: 14, Q8_0: 15

5. 验证

• GGUF 结构有效：True
• 真实推理测试（每个文件都由 llama.cpp 加载并生成一次文本）：yes

文件	llama.cpp 版本	可执行文件 SHA-256
Sakura-SmolLM2-135M-Instruct-BF16-Q4_K_M.gguf	b11669-620d2225c	3b7167843ffb0f3677be964dc80cfc59a0167c82e11eff62f6334f66e831f209

6. 下载与使用

llama-cli -m <file>.gguf -p "Hello" -n 32

需要支持该架构的 llama.cpp 版本。如果某个量化被拆分为带编号的多个部分，请将它们放在一起并加载第一个。

7. 局限

除非上文列出了实测数值，否则不声称任何编程基准、困惑度或质量分数。

8. 许可证信息

源许可证：apache-2.0。apache-2.0 allows redistribution with attribution

完整独立的中文版：README_zh.md