Download docs/development-artifacts/context-gradient-v1/report.md from zsqzz/llm-memory-editability: direct link, hf CLI and curl.
- Browser
- Download file 6.47 kB
-
https://huggingface.co/datasets/zsqzz/llm-memory-editability/resolve/main/docs/development-artifacts/context-gradient-v1/report.md
- Command line
-
hf download hf://datasets/zsqzz/llm-memory-editability/docs/development-artifacts/context-gradient-v1/report.md
-
curl -L -o report.md https://huggingface.co/datasets/zsqzz/llm-memory-editability/resolve/main/docs/development-artifacts/context-gradient-v1/report.md
上下文统计与真实早期梯度:完整开发结果
固定两开发世界×两初始化;4条共同中性短轨迹,0/32/128步,72个测量臂。所有统计均为开发描述;没有拟合预测比例、选取最佳状态或做显著性检验。
每个测量臂使用完整2048篇文档×10位置轮换。方向表先平均两个组织相对中性组织的差值,再平均初始化和世界;共8个配对对比,但只有两个开发世界。
全部预定状态与监督部分
| 步数 | 监督部分 | 预测器 | 平均余弦 | 最小–最大 | 平均相对误差 | 平均范数比 |
|---|---|---|---|---|---|---|
| 0 | answer | position | 0.2948 | 0.2853–0.3143 | 1.1557 | 0.9447 |
| 0 | answer | token | 0.2943 | 0.2850–0.3141 | 1.8327 | 1.8578 |
| 0 | answer | shuffled | 0.0030 | -0.0087–0.0149 | 1.0117 | 0.1567 |
| 0 | eos | position | 0.0022 | -0.0136–0.0275 | 1.0019 | 0.0647 |
| 0 | eos | token | 0.0080 | -0.0068–0.0267 | 1.0210 | 0.2139 |
| 0 | eos | shuffled | 0.0022 | -0.0136–0.0275 | 1.0019 | 0.0647 |
| 0 | combined | position | 0.2605 | 0.2485–0.2708 | 1.1094 | 0.8068 |
| 0 | combined | token | 0.2603 | 0.2486–0.2702 | 1.6470 | 1.5942 |
| 0 | combined | shuffled | 0.0089 | -0.0192–0.0580 | 1.0082 | 0.1375 |
| 32 | answer | position | 0.4173 | 0.4010–0.4315 | 2.3234 | 2.5550 |
| 32 | answer | token | 0.4166 | 0.4003–0.4311 | 4.6996 | 5.0270 |
| 32 | answer | shuffled | 0.0046 | -0.0085–0.0196 | 1.0844 | 0.4237 |
| 32 | eos | position | 0.0053 | -0.1618–0.1426 | 1.0351 | 0.2505 |
| 32 | eos | token | -0.0163 | -0.0624–0.0394 | 1.3114 | 0.8215 |
| 32 | eos | shuffled | 0.0053 | -0.1618–0.1426 | 1.0351 | 0.2505 |
| 32 | combined | position | 0.3179 | 0.2959–0.3397 | 2.6719 | 2.8151 |
| 32 | combined | token | 0.3172 | 0.2953–0.3391 | 5.3247 | 5.5563 |
| 32 | combined | shuffled | -0.0010 | -0.0368–0.0569 | 1.1100 | 0.4812 |
| 128 | answer | position | 0.3706 | 0.3373–0.3977 | 2.6423 | 2.8411 |
| 128 | answer | token | 0.3697 | 0.3373–0.3970 | 5.2860 | 5.5719 |
| 128 | answer | shuffled | 0.0048 | -0.0076–0.0156 | 1.1049 | 0.4714 |
| 128 | eos | position | 0.0521 | -0.1245–0.1760 | 1.2885 | 0.8605 |
| 128 | eos | token | 0.0389 | 0.0050–0.0896 | 2.9347 | 2.7934 |
| 128 | eos | shuffled | 0.0521 | -0.1245–0.1760 | 1.2885 | 0.8605 |
| 128 | combined | position | 0.3395 | 0.3097–0.3656 | 2.7691 | 2.9408 |
| 128 | combined | token | 0.3404 | 0.3121–0.3658 | 5.5045 | 5.7624 |
| 128 | combined | shuffled | -0.0221 | -0.0654–0.0316 | 1.1344 | 0.5091 |
相对误差为‖真实差值−预测差值‖/‖真实差值‖;范数比为‖预测差值‖/‖真实差值‖。没有用真实梯度估计最佳缩放。EOS目标恒定,shuffled在该部分与position相同,因此EOS的两者相等不算负对照失败,也不算关联预测证据。
主终点的全部配对块
| 世界 | 初始化 | 组织差值 | position余弦 | token余弦 | shuffled余弦 |
|---|---|---|---|---|---|
| 0 | 0 | company−neither | 0.2869 | 0.2869 | 0.0005 |
| 0 | 0 | project−neither | 0.2853 | 0.2850 | 0.0149 |
| 0 | 1 | company−neither | 0.3109 | 0.3108 | -0.0087 |
| 0 | 1 | project−neither | 0.3143 | 0.3141 | 0.0118 |
| 1 | 0 | company−neither | 0.2934 | 0.2925 | -0.0025 |
| 1 | 0 | project−neither | 0.2930 | 0.2928 | 0.0076 |
| 1 | 1 | company−neither | 0.2870 | 0.2861 | -0.0086 |
| 1 | 1 | project−neither | 0.2872 | 0.2866 | 0.0095 |
数据与数值控制
72/72臂完成,原始梯度与源码哈希复核通过。阻断跨事实注意力后,主投影梯度的最大相对消减残差为4.33e-08,低于运行前固定的1e-4阈值;不解释这些数值残差的余弦。
监督bigram计数在三组织间相同;城市×人物上下文统计在开放时不同,阻断后相同。这是对共同曝光和掩码的审计,不是表示或行为机制证明。
各模块的真实梯度差异
下表为开放注意力、答案部分、两个组织相对中性组织的差值;报告模块内范数,不能把不同参数块的欧氏范数解释为因果重要性。
| 步数 | 模块 | 平均差值范数 | 相对两臂平均范数 |
|---|---|---|---|
| 0 | token | 0.00982645 | 0.0200425 |
| 0 | position | 0.00140173 | 0.00940816 |
| 0 | blocks.0.attention.query | 0.000257498 | 0.0478393 |
| 0 | blocks.0.attention.key | 0.000263943 | 0.0501012 |
| 0 | blocks.0.attention.value | 0.00489436 | 0.0281848 |
| 0 | blocks.0.attention | 0.019808 | 0.0260856 |
| 0 | blocks.0.mlp | 0.0199909 | 0.0191931 |
| 0 | blocks.3.mlp | 0.0117847 | 0.0181113 |
| 0 | blocks.7.mlp | 0.00851532 | 0.0175424 |
| 32 | token | 0.00343339 | 0.00669271 |
| 32 | position | 0.000325903 | 0.0047403 |
| 32 | blocks.0.attention.query | 0.000102151 | 0.0339991 |
| 32 | blocks.0.attention.key | 0.000105125 | 0.0346381 |
| 32 | blocks.0.attention.value | 0.00181802 | 0.0235353 |
| 32 | blocks.0.attention | 0.00728407 | 0.0209926 |
| 32 | blocks.0.mlp | 0.00682808 | 0.0128873 |
| 32 | blocks.3.mlp | 0.00305384 | 0.00796859 |
| 32 | blocks.7.mlp | 0.00178565 | 0.00527191 |
| 128 | token | 0.00337177 | 0.00686375 |
| 128 | position | 0.000356374 | 0.00375431 |
| 128 | blocks.0.attention.query | 0.00011088 | 0.0251463 |
| 128 | blocks.0.attention.key | 0.00011178 | 0.0252903 |
| 128 | blocks.0.attention.value | 0.00173464 | 0.0276434 |
| 128 | blocks.0.attention | 0.00662339 | 0.0247392 |
| 128 | blocks.0.mlp | 0.00589619 | 0.0180513 |
| 128 | blocks.3.mlp | 0.00254916 | 0.0223958 |
| 128 | blocks.7.mlp | 0.00161888 | 0.0190356 |
attention无后缀的行只包含输出投影及其bias;Q/K/V分别列出。token是绑定的输入embedding/输出head,不能只解释为输出层。所有其他层、EOS、combined及company−project均保存在完整JSON/CSV中,不只保留上述示例行。
复现
PYTHONPATH=src python scripts/run_bios_context_gradient.py summarize
PYTHONPATH=src python scripts/report_bios_context_gradient.py
运行前定义与全部固定指标见快照和锁。原始产物位于results/bios-context-gradient-v1,包含12个共同参数/优化器状态、72臂的答案/EOS全模型梯度和三个预测器的主投影矩阵。首次GPU0加载触发ECC错误,无梯度产出;随后使用GPU6完成,设计及预算不变。
