Modification Summary (Trial 16)
Table with columns: Field, Value| Field | Value |
|---|
| Source model | lightx2v/MiniMax-H3-Prompt-Rewriter-Qwen3.8-27B-Ref2VA |
| Tool | Heretic v2.1.0, ARA modifier |
| Date | 2026-11-11 |
| Selection basis | Pareto frontier index 1 from a 30-run search |
| Export strategy | merge (correction baked into the weights, not an adapter) |
| Precision | BF16 (full BF16 weights reloaded from disk at merge time; 4-bit was used only during the search phase) |
| Weight files | 12 shards + model-mtp.safetensors = 13 files |
| Total weight bytes | 55,563,007,192 — bit-identical to the source model, no tensor loss |
Refusal & Fidelity Metrics
Table with columns: Metric, Source model, This model| Metric | Source model | This model |
|---|
| Refusal rate (KeywordRate) | 99 / 100 | 25 / 100 |
| KL divergence | 0 (by definition) | 0.0465 nats |
- Refusal rate: 100 prompts from
mlabonne/harmful_behaviors, generated one by one, refusal detected by keyword.
- KL divergence: 100 harmless prompts from
mlabonne/harmless_alpaca; full-vocabulary next-token distributions compared before/after, KL(source ‖ this model), averaged over prompts, in nats.
Modification Scope
ARA only touched two Linear module types within a specified layer range:
Table with columns: Parameter, Value, Meaning| Parameter | Value | Meaning |
|---|
start_layer_index | 1 | First layer (inclusive) |
end_layer_index | 44 | Last layer (exclusive) |
preserve_good_behavior_weight | 0.7683 | Constraint strength for normal behavior |
steer_bad_behavior_weight | 0.0001032 | Push away from refusals (at the search lower bound 0.0001) |
overcorrect_relative_weight | 1.0051 | Relative overcorrection amount |
neighbor_count | 2 | Number of reference neighbor samples |
- Modified: layers 1–43, 43 of 64 layers total — each layer's
attn.o_proj and mlp.down_proj, 86 modules in all.
- Untouched: layer 0, layers 44–63 (21 layers), token embeddings, all Norm layers, the vision tower (0.461B), and the MTP head (
model-mtp.safetensors, 0.425B / 810 MB).
Usage
This is a standard Transformers checkpoint, drop-in compatible with the source model. For usage instructions (inference examples, transformers / vLLM setup, etc.), please refer to the
source repository — just replace the model id with funnygeeker/MiniMax-H3-Prompt-Rewriter-Qwen3.8-27B-Ref2VA-heretic.
Usage Warnings
-
Risk of Sensitive or Controversial Outputs: This model's safety filtering has been significantly reduced, and it may generate sensitive, controversial, or inappropriate content. Users should exercise caution and rigorously review generated outputs.
-
Not Suitable for All Audiences: Due to limited content filtering, the model's outputs may be inappropriate for public settings, underage users, or applications requiring high security.
-
Legal and Ethical Responsibilities: Users must ensure their usage complies with local laws and ethical standards. Generated content may carry legal or ethical risks, and users are solely responsible for any consequences.
-
Research and Experimental Use: It is recommended to use this model for research, testing, or controlled environments, avoiding direct use in production or public-facing commercial applications.
-
Monitoring and Review Recommendations: Users are strongly advised to monitor model outputs in real time and conduct manual reviews when necessary to prevent the dissemination of inappropriate content.
-
No Default Safety Guarantees: Unlike standard models, this model has not undergone rigorous safety optimization. The author bears no responsibility for any consequences arising from its use.
My quantized (oQ / MLX) builds of this model:
中文
本模型是 lightx2v/MiniMax-H3-Prompt-Rewriter-Qwen3.8-27B-Ref2VA 的去审查(decensored)版本,使用 Heretic v2.1.0(ARA 修改器)制作。完整 BF16 权重检查点,不是量化产物也不是 adapter:修正量已焊进权重,总字节数与源模型逐位相等(55,563,007,192 字节)。
改动摘要(Trial 16)
Table with columns: 字段, 值| 字段 | 值 |
|---|
| 源模型 | lightx2v/MiniMax-H3-Prompt-Rewriter-Qwen3.8-27B-Ref2VA |
| 工具 | Heretic v2.1.0,ARA 修改器 |
| 日期 | 2026-11-11 |
| 选择依据 | 30 轮搜索的 Pareto 前沿 index 1 |
| 导出方式 | merge(修正量焊进权重,非 adapter) |
| 精度 | BF16(合并时从磁盘重载完整 BF16 原权重;4-bit 仅用于搜索阶段) |
| 权重文件 | 12 个分片 + model-mtp.safetensors,共 13 个 |
| 权重总字节 | 55,563,007,192 —— 与源模型逐位相等,无张量丢失 |
拒绝率与保真度指标
Table with columns: 指标, 源模型, 本模型| 指标 | 源模型 | 本模型 |
|---|
| 拒绝率(KeywordRate) | 99 / 100 | 25 / 100 |
| KL 散度 | 0(按定义) | 0.0465 nats |
- 拒绝率:对
mlabonne/harmful_behaviors 的 100 条提示词逐条生成,用关键词判定是否拒绝。
- KL 散度:对
mlabonne/harmless_alpaca 的 100 条无害提示词,比较改前/改后的下一个 token 全词表分布,KL(源模型 ‖ 本模型),对提示词取平均,单位 nats。
改动范围
ARA 只在指定层区间内改动两类 Linear 模块:
Table with columns: 参数, 值, 含义| 参数 | 值 | 含义 |
|---|
start_layer_index | 1 | 起始层(含) |
end_layer_index | 44 | 结束层(不含) |
preserve_good_behavior_weight | 0.7683 | 保留正常行为的约束强度 |
steer_bad_behavior_weight | 0.0001032 | 远离拒绝行为的推力(已贴到搜索下界 0.0001) |
overcorrect_relative_weight | 1.0051 | 相对过校正量 |
neighbor_count | 2 | 参考的邻近样本数 |
- 实际改动:第 1–43 层,共 43 / 64 层,每层的
attn.o_proj 与 mlp.down_proj,合计 86 个模块。
- 未被改动:第 0 层、第 44–63 层(共 21 层)、词嵌入、所有 Norm、视觉塔(0.461B)、MTP 头(
model-mtp.safetensors,0.425B / 810 MB)。
使用方式
标准 Transformers 检查点,与源模型完全兼容,可即插即用。使用方式(推理示例、transformers / vLLM 配置等)请参考
源仓库,把模型 ID 替换为 funnygeeker/MiniMax-H3-Prompt-Rewriter-Qwen3.8-27B-Ref2VA-heretic 即可。
使用警告
-
敏感或有争议输出的风险:本模型的安全过滤已被大幅削弱,可能生成敏感、有争议或不适宜的内容。用户应谨慎对待并严格审查生成内容。
-
不适合所有受众:由于内容过滤有限,模型输出可能不适合公开场合、未成年用户或高安全要求的应用场景。
-
法律与道德责任:用户必须确保其使用方式符合当地法律与道德标准。生成内容可能带来法律或道德风险,用户自行承担由此产生的任何后果。
-
研究与实验用途:建议将本模型用于研究、测试或受控环境,避免直接用于生产或面向公众的商业应用。
-
监控与审查建议:强烈建议用户实时监控模型输出,必要时进行人工审查,以防不当内容传播。
-
无默认安全保证:与标准模型不同,本模型未经过严格的安全优化。作者对本模型使用产生的任何后果不承担任何责任。
相关产物
本模型的我自制的量化(oQ / MLX)版本: