Files
WpyQwq 643e22ecb9 Natural Memory NM2.1: 记忆路由器分叉、数据集缺陷修复与全轴评测证据
- 引入 MemoryRouterXL 与 v5/v6 流式多线程训练/编码管线
- 修复 prepare_memory_router_dataset 候选池重建缺陷(mega 家族 3568x 加速,输出逐字节相同)
- 修复 v5 被破坏的拒答与多跳标签(train 未知样本 319 -> 16319,multi_hop 平均正例 1.00 -> 2.00)
- 同存储预算下 V2-128 v6 逐轴 22/22 通过:Top-1 41.12% -> 94.62%,未知拒答 0.00% -> 100.00%
- 记录三条被实测推翻的显然优化(logits_to_keep=1 反而慢 55%、XL 容量未带来收益)
- 记忆手术跨架构可移植性 14/14,读写关闭时与原生模型逐位相同
2026-09-19 11:11:31 +08:00

17 lines
1.6 KiB
Markdown

| model | model_type | hidden | layers | container | 2nd positional | identity | surgery |
|---|---|---|---|---|---|---|---|
| tiny-random-LlamaForCausalLM | llama | 16 | 2 | model.layers | OTHER | identical | ok (no shims) |
| tiny-random-MistralForCausalLM | mistral | 32 | 2 | model.layers | OTHER | identical | ok (no shims) |
| tiny-Qwen2ForCausalLM-2.5 | qwen2 | 8 | 2 | model.layers | OTHER | identical | ok (no shims) |
| tiny-Qwen3ForCausalLM | qwen3 | 8 | 2 | model.layers | OTHER | identical | ok (no shims) |
| tiny-random-Gemma2ForCausalLM | gemma2 | 32 | 1 | model.layers | position_embeddings | identical | ok (no shims) |
| tiny-random-Gemma3ForCausalLM | gemma3_text | 16 | 2 | model.layers | position_embeddings | identical | ok (no shims) |
| tiny-random-Starcoder2ForCausalLM | starcoder2 | 32 | 2 | model.layers | OTHER | identical | ok (no shims) |
| tiny-random-OlmoeForCausalLM | olmoe | 64 | 2 | model.layers | OTHER | identical | ok (no shims) |
| tiny-random-GraniteMoeForCausalLM | granitemoe | 32 | 2 | model.layers | OTHER | identical | ok (no shims) |
| tiny-random-MixtralForCausalLM | mixtral | 64 | 2 | model.layers | position_embeddings | identical | ok (no shims) |
| tiny-random-CohereForCausalLM | cohere | 32 | 2 | model.layers | OTHER | identical | ok (no shims) |
| tiny-random-PhiForCausalLM | phi | 32 | 2 | model.layers | OTHER | identical | ok (no shims) |
| tiny-random-OPTForCausalLM | opt | 16 | 5 | model.decoder.layers | OTHER | identical | ok (no shims) |
| tiny-random-BartForCausalLM | bart | 16 | 2 | model.decoder.layers | OTHER | identical | ok (no shims) |