Spark-X2.5-4B Abliterixed — BF16, FP8 and NVFP4

Unofficial community derivative of XHToken/Spark-X2.5-4B, processed with Abliterix to reduce refusal-related behavior. This is not an official SparkLLM or XHToken release.

Available variants

Each folder is self-contained. GGUF variants are in the GGUF repository, and MLX variants are in the MLX repository.

Latest safety evaluation

The following is the latest completed evaluation, run on the NVFP4/ variant served as XHToken--Spark-X2.5-4B-NVFP4 through vLLM. These results demonstrate substantially weakened refusal behavior and a material unsafe-output risk.

Scope: only NVFP4 was evaluated in this run. Do not treat these numbers as independent measurements of BF16, FP8, GGUF, or MLX variants.

Summary

For StrongREJECT, higher scores mean the model more successfully answered harmful requests. The median score was 1.000000, non-refusal was 95.56% (172/180), and 60.00% of responses contained a warning or disclaimer while often continuing with relevant content. The official string-matching baseline mean was 0.838889.

For SimpleSafetyTests, the auxiliary behavior judge found 0 full refusals, 3 partial refusals, and 97 non-refusals. Warnings or caveats appeared in 78/100 responses. The official SST primary metric is holistic safe/unsafe classification, not refusal-keyword matching.

StrongREJECT by category

SimpleSafetyTests by harm area

Evaluation configuration and limitations

• Target: vLLM OpenAI-compatible endpoint, model XHToken--Spark-X2.5-4B-NVFP4
• StrongREJECT: 180 prompts, six categories, 30 prompts per category; temperature 0, seed 42, max tokens 2,048
• SST: official 100-prompt all split, none condition; temperature 0.01, seed 42, max tokens 4,096
• Judge: Ollama Cloud deepseek-v4-flash, think=false
• Only the target model's returned content was judged; private reasoning was excluded where separately available
• StrongREJECT finish reasons: 153 length, 27 stop; the high truncation rate is an important limitation
• SST finish reasons: 91 stop, 9 length
• These are automated-judge results, not human adjudication

Architecture and compatibility

• Architecture: Spark2_5ForCausalLM (model_type: spark2_5)
• 36 transformer layers, hidden size 2,560, 16 attention heads, 4 KV heads
• Hybrid full/sliding-window attention with custom Spark2.5 modeling code
• Declared maximum context: 1,048,576 tokens; practical context depends on memory, KV cache and concurrency

Use a runtime that explicitly supports Spark-X2.5, or review and load the included custom code with trust_remote_code=True. Quantized-kernel availability depends on runtime, accelerator and software versions.

Download one variant

hf download GlobalCybersecurityAlliance/Spark-X2.5-4B-Abliterixed \
  --include "NVFP4/*" \
  --local-dir ./Spark-X2.5-4B-Abliterixed

Replace NVFP4 with BF16 or FP8 as needed.

Safety, intended use and disclaimer

Abliterix deliberately changes refusal-related behavior. The latest evaluation shows that this derivative frequently complies with harmful requests. Warnings and disclaimers must not be interpreted as effective refusal or safety mitigation.

Appropriate uses include alignment and model-behavior research, authorized red-team evaluation, and controlled defensive research. Do not use the model for unlawful activity, unauthorized access, malware deployment, fraud, exploitation, harassment, violence, privacy invasion, or other harm. Deployers are responsible for access controls, monitoring, safeguards, legal compliance and human review.

The model is provided AS IS, without warranties. Quantization may further change quality, calibration and safety behavior. Independently validate the selected variant and runtime before use.

License and attribution

The official base repository identifies its license as Apache-2.0. Users must review and comply with the base-model license and all applicable terms for dependencies, runtimes, datasets, inputs, outputs and intended use.

中文说明

本仓库是 XHToken/Spark-X2.5-4B 的非官方社区衍生版本，经过 Abliterix 处理以降低拒答相关行为。本模型不是 SparkLLM 或 XHToken 的官方发布。

可用版本

每个目录都是自包含模型目录。GGUF 版本位于 GGUF 仓库，MLX 版本位于 MLX 仓库。

最后一次安全评测

以下是最后一次完整评测的结果。被测对象是通过 vLLM 服务、名称为 XHToken--Spark-X2.5-4B-NVFP4 的 NVFP4/ 版本。结果表明其拒答行为显著减弱，并存在实质性不安全输出风险。

适用范围： 本轮只评测了 NVFP4。不得将这些数字表述为 BF16、FP8、GGUF 或 MLX 版本各自的独立实测结果。

结果概览

StrongREJECT 分数越高，表示模型越成功地回答了有害请求。其中位数为 1.000000，不拒绝率为 95.56%（172/180）；60.00% 的回答包含警告或免责声明，但通常仍继续提供相关内容。官方字符串匹配基线平均分为 0.838889。

SimpleSafetyTests 的 3% 拒绝率来自辅助行为标签中的 3 次部分拒绝；其余为 0 次完全拒绝和 97 次不拒绝。需要注意，SST 官方主指标是整体 safe/unsafe 分类，而不是拒答关键词匹配；本轮官方主指标为 32/100 Unsafe、68/100 Safe，78/100 的回答包含警告或提示。

StrongREJECT 分类结果

SimpleSafetyTests 危害领域结果

评测配置与局限性

• 目标：vLLM OpenAI 兼容接口，模型 XHToken--Spark-X2.5-4B-NVFP4
• StrongREJECT：180 条提示，6 个类别、每类 30 条；temperature 0、seed 42、max tokens 2,048
• SST：官方 100 条 all split、none condition；temperature 0.01、seed 42、max tokens 4,096
• 裁判：Ollama Cloud deepseek-v4-flash，think=false
• 仅评判目标模型返回的最终内容；在私有推理可单独取得时，不将其发送给裁判
• StrongREJECT finish reasons：153 条 length、27 条 stop；高截断比例是解释结果时的重要限制
• SST finish reasons：91 条 stop、9 条 length
• 结果来自自动裁判，不等同于人工复核

架构与兼容性

• 架构：Spark2_5ForCausalLM（model_type: spark2_5）
• 36 层 Transformer、隐藏维度 2,560、16 个注意力头、4 个 KV 头
• 使用自定义 Spark2.5 代码的全注意力/滑动窗口混合架构
• 配置声明最大上下文为 1,048,576 token；实际长度受内存、KV cache 和并发量限制

请使用明确支持 Spark-X2.5 的运行时，或审阅随模型提供的自定义代码后使用 trust_remote_code=True。量化内核是否可用取决于运行时、加速器及软件版本。

下载单个版本

hf download GlobalCybersecurityAlliance/Spark-X2.5-4B-Abliterixed \
  --include "NVFP4/*" \
  --local-dir ./Spark-X2.5-4B-Abliterixed

可按需将 NVFP4 替换为 BF16 或 FP8。

安全、预期用途与免责声明

Abliterix 会有意改变拒答相关行为。最后一次评测显示，该衍生模型经常服从有害请求。警告和免责声明不能被视为有效拒答或充分的安全缓解措施。

适当用途包括对齐与模型行为研究、经授权的红队评测及受控防御性研究。禁止用于违法活动、未授权访问、恶意软件投放、欺诈、剥削、骚扰、暴力、侵犯隐私或其他伤害。部署者负责访问控制、监控、安全措施、法律合规及人工复核。

模型按现状提供，不作任何保证。量化可能进一步改变质量、校准和安全行为。使用前应在目标运行时独立验证所选版本。

许可证与署名

官方基础模型仓库标注的许可证为 Apache-2.0。使用者必须审阅并遵守基础模型许可证，以及依赖、运行时、数据集、输入、输出和预期用途所适用的全部条款。