Sakura Qwen3.8-27B Humanlike Chat 2.0 JEV — Uncensored AutoRound Q3_K_M GGUF

Sakura Qwen3.8-27B Humanlike Chat 2.0 JEV combines the Humanlike Chat 2.0 / uncensored lineage with JEV decision capabilities and is distributed as an AutoRound-optimized Q3_K_M GGUF package.

It provides two complementary usage paths:

System Two: normal generative chat/reasoning/coding
System One: calibrated decision/judgment outputs through the bundled JEV head

Model Provenance

Qwen/Qwen3.8-27B
  ↓
huihui-ai/Huihui-Qwen3.8-27B-abliterated
  ↓
LessThanThreeAI/Qwen3.8-27B-Humanlike-Chat-2.0
  +
autotrust/JEV-27B
  ↓
Sakura AutoRound Q3_K_M GGUF

Base Lineage (Chat & Reasoning): LessThanThreeAI/Qwen3.8-27B-Humanlike-Chat-2.0
JEV Decision Source: autotrust/JEV-27B (Upstream commit revision: 51740a8891c2a8baefd969237fd44187b3e3a115)

Package Architecture & Two-System Design

This release is a model package supporting two distinct operational modalities:

1. System Two: Generative Conversation

Standard generative chat, programming, and reasoning running on stock GGUF runtimes (llama.cpp, Ollama, etc.).

Qualitative Release Validation (Smoke Tests):

German Chat: Precise thermodynamic definitions, coherent phrasing.
English Chat: Concise, actionable software engineering practices.
Coding: Clean, idiomatic Python implementations (is_palindrome).
Structured JSON: Strict JSON objects without unwanted markdown formatting.
Humanlike Behavior: Natural, conversational tone without formulaic assistant cliches or leaking internal decision control tokens.
Uncensored / Abliterated Behavior: Objective technical explanations of physical lock mechanics without synthetic refusals.

(Note: These validations represent smoke tests for qualitative release verification, not an exhaustive general benchmark.)

2. System One: Calibrated Decision & Judgment

Fast System-1 judgment path utilizing the bundled decision projection head and calibration metadata in an OpenJEV-compatible runtime.

Bundled Decision Components:

head.safetensors: Decision projection head tensor with shape [24, 5120] plus bias [24].
calibration.json: Slot-specific temperature scaling constants:
- $T_{\text{noul}} = 1.014313$
- $T_{\text{choice}} = 1.016128$
- $T_{\text{score}} = 1.003621$
decision_head.json: Architectural projection mappings.
judge_config.json: Routing judge parameters.
chat_template.jinja: Jinja chat template with System-One token sequences.

Validated Decision Modalities:

noul (Slots [0:2], 2 options)
score (Slots [2:8], 6 discrete score ratings 0–5)
choice (Slots [8:24], tested across 2, 4, 8, and 16 candidates)
Tool / Function Decision (Routing calls to tools)

Quantization & Technical Verification

Quantized using AutoRound GPU Q3 (Q3_K_M):

Optimization Parameters: 64 / 64 transformer layers optimized, 200 iterations per block, 64 calibration samples, sequence length 512.
Hardware & Stack: AMD Radeon 8060S (gfx1151), ROCm 7.13 nightly.
Export Format: Q3_K_M GGUF.

Verification Metrics & Parity Analysis

PEFT vs. Merged HF Spot-Check: 4 / 4 representative spot-check cases achieved 100% top-choice match. Maximum observed numerical deviation: 0.001878 (consistent with expected numerical precision differences).
System-One Release-Gate Verification:
- 100% top-choice agreement on the final System-One release-gate / smoke set.
- On the final System-One release-gate cases: Mean KL divergence $< 0.001$, JS divergence $< 0.0005$, Probability drift $< 0.002$.
Original JEV Reference Comparison:
- Overall agreement with the original Qwen+JEV reference suite is 84.47%, with 40 / 40 (100%) agreement on the Tool / Function calling subset.
- Context: The 84.47% agreement with the original JEV reference characterizes the Humanlike-JEV merge itself rather than degradation caused by the final AutoRound Q3_K_M quantization. The differences concentrate on subjective and conversational moderation scenarios where Humanlike conversational fine-tuning deliberately diverges from vanilla Qwen.

File Manifest & Checksums

Filename	Size (Bytes)	Size (GiB)	Description
`Sakura-Humanlike-Chat-2.0-JEV-Q3_K_M.gguf`	13,500,736,864	12.57 GiB	AutoRound Q3_K_M quantized backbone
`mmproj-model.gguf`	1,842,940,128	1.72 GiB	Multimodal vision projector
`head.safetensors`	491,768	< 1 MiB	OpenJEV System-1 decision head `[24, 5120]`
`calibration.json`	999	< 1 KiB	Temperature calibration constants
`decision_head.json`	477	< 1 KiB	Decision head configuration
`judge_config.json`	1,431	< 2 KiB	Evaluation & routing judge config
`chat_template.jinja`	8,952	~9 KiB	Jinja chat template with System-One support
`SHA256SUMS`	620	< 1 KiB	Verification checksums

Primary Checksum:

Sakura-Humanlike-Chat-2.0-JEV-Q3_K_M.gguf:
ef032c3fd92ba106e291c27dbd1e02e07b308ec6bfebdffc0d88e12804c45d0f

Usage Instructions

System Two (Interactive Chat via llama.cpp)

llama-cli -m Sakura-Humanlike-Chat-2.0-JEV-Q3_K_M.gguf \
  -p "<|im_start|>user\nErkläre kurz das Prinzip von Entropie.<|im_end|>\n<|im_start|>assistant\n" \
  -n 512 --temp 0.7 -ngl 99

System One (OpenJEV Runtime / Python)

Load the pre-merged backbone with the companion head.safetensors and calibration.json:

import json
import torch
from safetensors import safe_open

with safe_open("head.safetensors", framework="pt") as f:
    head_w = f.get_tensor("proj.weight") # [24, 5120]
    head_b = f.get_tensor("proj.bias")   # [24]

with open("calibration.json", "r", encoding="utf-8") as f:
    calib = json.load(f)["per_kind"]

Related Sakura models

webmp3/Sakura-Humanlike-Chat-2.0-AutoRound-GGUF: Base Humanlike / uncensored AutoRound Q3_K_M release without the JEV decision package.

License & Credits

Base weights derived from Qwen 3.8 / Humanlike-Chat-2.0 under Apache 2.0.
System-1 Decision Architecture developed by OpenJEV (autotrust/JEV-27B).
Quantized using Intel AutoRound on AMD Radeon ROCm hardware.

中文说明 · 樱花 (Simplified Chinese)

English above. 本节为上文的中文翻译(Sakura = 樱花 yīnghuā);完整的独立中文版见 README_zh.md。

Sakura Qwen3.8-27B Humanlike Chat 2.0 JEV — Uncensored AutoRound Q3_K_M GGUF

Sakura Qwen3.8-27B Humanlike Chat 2.0 JEV 把 Humanlike Chat 2.0 / uncensored 的谱系与 JEV 决策能力结合在一起,以经过 AutoRound 优化的 Q3_K_M GGUF 包的形式发布。

它提供两条互补的使用路径:

System Two: 常规的生成式聊天/推理/编程
System One: 通过随附的 JEV 头给出经过校准的决策/判断输出

模型谱系

Qwen/Qwen3.8-27B
  ↓
huihui-ai/Huihui-Qwen3.8-27B-abliterated
  ↓
LessThanThreeAI/Qwen3.8-27B-Humanlike-Chat-2.0
  +
autotrust/JEV-27B
  ↓
Sakura AutoRound Q3_K_M GGUF

基础谱系(聊天与推理): LessThanThreeAI/Qwen3.8-27B-Humanlike-Chat-2.0
JEV 决策来源: autotrust/JEV-27B(上游提交修订版本:51740a8891c2a8baefd969237fd44187b3e3a115)

包的架构与双系统设计

本次发布是一个 模型包,支持两种不同的运行方式:

1. System Two:生成式对话

标准的生成式聊天、编程和推理,可在原版 GGUF 运行时(llama.cpp、Ollama 等)上运行。

定性的发布验证(冒烟测试):

德语聊天: 热力学定义准确,措辞连贯。
英语聊天: 简明、可操作的软件工程实践。
编程: 干净、地道的 Python 实现(is_palindrome)。
结构化 JSON: 严格的 JSON 对象,没有多余的 markdown 格式。
Humanlike 行为: 自然的对话语气,没有公式化的助手套话,也没有泄漏内部决策控制 token。
Uncensored / Abliterated 行为: 对物理锁机构给出客观的技术解释,没有人为的拒绝。

(说明:这些验证是用于定性发布核查的冒烟测试,不是详尽的通用基准测试。)

2. System One:经过校准的决策与判断

快速的 System-1 判断路径,在兼容 OpenJEV 的运行时中使用随附的决策投影头和校准元数据。

随附的决策组件:

head.safetensors:决策投影头张量,形状为 [24, 5120],另有偏置 [24]。
calibration.json:各槽位的温度缩放常数:
- $T_{\text{noul}} = 1.014313$
- $T_{\text{choice}} = 1.016128$
- $T_{\text{score}} = 1.003621$
decision_head.json:架构投影映射。
judge_config.json:路由评判参数。
chat_template.jinja:带有 System-One token 序列的 Jinja 聊天模板。

已验证的决策形式:

noul(槽位 [0:2],2 个选项)
score(槽位 [2:8],6 个离散评分 0–5)
choice(槽位 [8:24],已在 2、4、8 和 16 个候选上测试)
Tool / Function Decision(把调用路由到工具)

量化与技术验证

使用 AutoRound GPU Q3(Q3_K_M)量化:

优化参数: 优化了 64 / 64 个 Transformer 层,每个块 200 次迭代,64 个校准样本,序列长度 512。
硬件与软件栈: AMD Radeon 8060S(gfx1151),ROCm 7.13 nightly。
导出格式: Q3_K_M GGUF。

验证指标与一致性分析

PEFT 对合并后 HF 的抽查: 4 / 4 个有代表性的抽查用例,首选项 100% 一致。观察到的最大数值偏差:0.001878(与预期的数值精度差异一致)。
System-One 发布关卡验证:
- 在最终的 System-One 发布关卡 / 冒烟测试集上,首选项一致率 100%。
- 在最终的 System-One 发布关卡用例上:平均 KL 散度 $< 0.001$,JS 散度 $< 0.0005$,概率漂移 $< 0.002$。
与原始 JEV 参照的比较:
- 与原始 Qwen+JEV 参照套件的总体一致率为 84.47%,其中工具 / 函数调用子集的一致率为 40 / 40(100%)。
- 背景: 与原始 JEV 参照的 84.47% 一致率刻画的是 Humanlike-JEV 合并本身,而不是最终 AutoRound Q3_K_M 量化造成的退化。差异集中在主观的、对话式的审核场景,在这些场景中 Humanlike 对话微调有意地偏离了原版 Qwen。

文件清单与校验和

文件名	大小(字节)	大小(GiB)	说明
`Sakura-Humanlike-Chat-2.0-JEV-Q3_K_M.gguf`	13,500,736,864	12.57 GiB	AutoRound Q3_K_M 量化的主干
`mmproj-model.gguf`	1,842,940,128	1.72 GiB	多模态视觉投影器
`head.safetensors`	491,768	< 1 MiB	OpenJEV System-1 决策头 `[24, 5120]`
`calibration.json`	999	< 1 KiB	温度校准常数
`decision_head.json`	477	< 1 KiB	决策头配置
`judge_config.json`	1,431	< 2 KiB	评测与路由评判配置
`chat_template.jinja`	8,952	~9 KiB	支持 System-One 的 Jinja 聊天模板
`SHA256SUMS`	620	< 1 KiB	校验和

主要校验和:

Sakura-Humanlike-Chat-2.0-JEV-Q3_K_M.gguf:
ef032c3fd92ba106e291c27dbd1e02e07b308ec6bfebdffc0d88e12804c45d0f

使用说明

System Two(通过 llama.cpp 交互式聊天)

llama-cli -m Sakura-Humanlike-Chat-2.0-JEV-Q3_K_M.gguf \
  -p "<|im_start|>user\nErkläre kurz das Prinzip von Entropie.<|im_end|>\n<|im_start|>assistant\n" \
  -n 512 --temp 0.7 -ngl 99

System One(OpenJEV 运行时 / Python)

将预先合并的主干与配套的 head.safetensors 和 calibration.json 一起加载:

import json
import torch
from safetensors import safe_open

with safe_open("head.safetensors", framework="pt") as f:
    head_w = f.get_tensor("proj.weight") # [24, 5120]
    head_b = f.get_tensor("proj.bias")   # [24]

with open("calibration.json", "r", encoding="utf-8") as f:
    calib = json.load(f)["per_kind"]

许可证与致谢

基础权重派生自 Qwen 3.8 / Humanlike-Chat-2.0,采用 Apache 2.0。
System-1 决策架构由 OpenJEV 开发(autotrust/JEV-27B)。
使用 Intel AutoRound 在 AMD Radeon ROCm 硬件上量化。

webmp3/Sakura-Humanlike-Chat-2.0-JEV-AutoRound-GGUF

Sakura Qwen3.8-27B Humanlike Chat 2.0 JEV — Uncensored AutoRound Q3_K_M GGUF

Model Provenance

Package Architecture & Two-System Design

1. System Two: Generative Conversation

2. System One: Calibrated Decision & Judgment

Quantization & Technical Verification

Verification Metrics & Parity Analysis

File Manifest & Checksums

Usage Instructions

System Two (Interactive Chat via llama.cpp)

System One (OpenJEV Runtime / Python)

Related Sakura models

License & Credits

Sakura Qwen3.8-27B Humanlike Chat 2.0 JEV — Uncensored AutoRound Q3_K_M GGUF

模型谱系

包的架构与双系统设计

1. System Two:生成式对话

2. System One:经过校准的决策与判断

量化与技术验证

验证指标与一致性分析

文件清单与校验和

使用说明

System Two(通过 llama.cpp 交互式聊天)

System One(OpenJEV 运行时 / Python)

相关的 Sakura 模型

许可证与致谢