← back to catalog · registered 2026-11-18 11:04

BG4JEC/Xiaoheimao-Qwen3.8-27B-abliterated

BG4JEC 27B GGUF second-order
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/BG4JEC%2FXiaoheimao-Qwen3.8-27B-abliterated"
Response includes
  • classification m8
  • files 3
  • author_summary 1 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M8
Primary method

Repackaging (quantization)

Applied on top of direct removal inherited from the base model.
Confidence
MEDIUM
Inherited from base model
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=1
  • assume M1 (base ablation) + M8 (GGUF quant) - default when producer unknown
Refusal direction extraction

No specific extraction method could be identified for this model. The producer either did not document it or used a proprietary pipeline.

What is a refusal direction? →
Downloads · 30-day
217
Likes
0
Model age
3d ago
created 2026-11-18
Downloads over time
Now 217 → from 0 ↑ 0%
080159239 0 on Nov 18 217 on Nov 22 Nov
Nov 18 → Nov 22 · 5 snapshots · spans 4 days

Genealogy 0 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Metadata

License
apache-2.0
Tags
safetensors gguf qwen3.8 27b lora iq4_xs roleplay uncensored base_model:huihui-ai/Huihui-Qwen3.8-27B-abliterated base_model:adapter:huihui-ai/Huihui-Qwen3.8-27B-abliterated license:apache-2.0 endpoints_compatible

Related

Total size
0 B
Files
3
Quantizations
1
Registered
2026-11-18 11:04
Last updated on HF
2026-11-18 12:00

Files by quantization

Auxiliary files 3 files 6.38 KB
README.md 4.37 KB 520cb71c download
.gitattributes 1.54 KB 99de8ec3 download
.gitignore 484 B 6d104c16 download

README current version from Hugging Face


license: apache-2.0
base_model:

  • huihui-ai/Huihui-Qwen3.8-27B-abliterated
    tags:
  • qwen3.8
  • 27b
  • lora
  • gguf
  • iq4_xs
  • roleplay
  • uncensored

Xiaoheimao-Qwen3.8-27B-abliterated

基于 huihui-ai/Huihui-Qwen3.8-27B-abliterated 后训练的群聊风格对话模型,融合了非 AI 的 Repeater 关键词联想金句引擎,并量化导出为可直接运行在单张 16GB Tesla V100 上的 IQ4_XS GGUF。

详细训练日志、5 维度盲测统计与对比样例见 TRAINING_AND_REPEATER_REPORT.md。


1. 模型训练与盲测表现

  • 训练配置:Google Colab NVIDIA A100-SXM4-80GB,全精度 BF16 LoRA(r=32, lora_alpha=64,可训练参数 0.8025%),19,130 条高密度群聊 + Repeater 金句融合样本(598 步,MAX_SEQ_LEN=512,全程关闭 <think>)。
  • Loss 收敛:基线验证集 eval_loss = 5.6527 → 最终 eval_loss = 3.9568(train_loss = 3.8977)。
  • 100 题零泄漏 5 维度盲测(对比真人群友基准 P50=8.0 字、P90=18.0 字、黑话命中 21.0%):
    • pure_llm:P50=7.0 字、P90=15.0 字、AI 说教率 0.0%、多行幻觉 0.0%、去重率 100%、群黑话命中 31.0%。
    • repeater_rag:P50=7.0 字、P90=14.0 字、AI 说教率 0.0%、多行幻觉 0.0%、照抄提示率 0.0%、群黑话命中 32.0%。

2. 16GB Tesla V100 部署指南

2.1 启动 llama-server

当前成品 Xiaoheimao-Qwen3.8-27B-abliterated-IQ4_XS.gguf 实测大小为 14.48 GiB(14,814.35 MiB,保留了完整词表与高精度层)。在 16GB V100 上请务必固定单槽位并使用 q8_0 KV Cache,同时关闭 Thinking 模式:

llama-server \
  -m Xiaoheimao-Qwen3.8-27B-abliterated-IQ4_XS.gguf \
  --alias xiaoheimao \
  -ngl 99 \
  -fa on \
  -ctk q8_0 -ctv q8_0 \
  -c 4096 -np 1 -b 512 \
  --jinja \
  --chat-template-kwargs '{"enable_thinking":false}' \
  --host 0.0.0.0 --port 8080

(若要在 16GB V100 上开到 16K–32K 超长上下文并腾出更多显存余量,可直接用本地 llama-quantize --allow-requantize 二次压成 ~12.5 GiB 的 IQ3_M 或 Q3_K_M:)

llama-quantize --allow-requantize \
  Xiaoheimao-Qwen3.8-27B-abliterated-IQ4_XS.gguf \
  Xiaoheimao-Qwen3.8-27B-abliterated-IQ3_M.gguf IQ3_M

2.2 挂接 Repeater 非 AI 联想金句与混合推理器

项目内置了移植自 Pallas-Bot repeater 插件的轻量级非 AI 金句引擎(inference/xiaoheimao_responder.py + outputs/repeater_memory_834513952.json),单次检索耗时仅 0.11–0.20 ms:

  • pure_llm:纯微调模型直接接话。
  • repeater_rag:根据当前消息的 jieba Top-2 关键词与近期话题滑动队列 _recent_topics 实时召回 2–4 条关联群金句写入 system 提示词,引导模型自然化用黑话(实测 0% 原样照抄)。
  • hybrid_intermittent:混合间歇触发模式——遇到群体复读浪潮时 0 Token 秒跟复读;命中高赞金句(count >= 5)时按 15% 概率直接空投原声金句;其余情况走 repeater_rag 生成,并按概率将带逗号的回复拆成多条短气泡发出(支持冷场主动破冰 cold_group_speak)。
# 安装轻量依赖
pip install jieba pypinyin openai

# 运行单轮测试
python3 inference/xiaoheimao_responder.py \
  --memory outputs/repeater_memory_834513952.json \
  --base-url http://127.0.0.1:8080/v1 \
  --mode hybrid_intermittent \
  --prompt "雷姐今天怎么又鸽了"

3. 仓库目录说明

README history 1 version

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-11-18 Upload Xiaoheimao-Qwen3.8-27B-abliterated code, docs, and LoRA adapter 1092cbc 4.4 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration