# AWDE 0620 CHSIMS FDE Token-Residual Results

日期：2026-06-20

## 实验目标

把 0619-FDE 中 MOSEI 表现较好的 `fde_tokenres_h160_a12_lr5e5_b8_d15` 迁移到 CHSIMS English 数据集上验证。为了看 seed 稳定性，本次使用 4 卡跑同一超参的 4 个 seed。

## 运行配置

- 代码目录：`/root/AWDE/0620chsims`
- 输出目录：`/root/siton-data-531cb60d91bd4013b805b412b0be2176/tlw/store/AWDE/0620chsims`
- 数据 pkl：`/root/siton-data-531cb60d91bd4013b805b412b0be2176/tlw/store/pkl/0617-CHSIMS-eng/chsims_awde_0617_eng_encoder_raw512_fp16.pkl`
- 启动脚本：`/root/AWDE/0620chsims/scripts/start_0620_chsims_fde_npu4.sh`
- 汇总脚本：`/root/AWDE/0620chsims/scripts/summarize_0620_chsims_fde_runs.py`

关键超参：

```text
hidden_dim=160
batch_size=8
lr=5e-5
dropout=0.15
fd_enhance_mode=token_residual
fd_token_alpha=0.12
desc_gate_mode=none
desc_alpha=0.0
temporal_align_type=eats
temporal_desc_bias=0.0
smooth_l1_beta=0.25
use_ema=true
selection_metric=composite
early_stop_patience=25
```

## 4-run 结果

| Run | Seed | Best | Source | Composite | Valid Acc-2 | Valid MAE | Valid Corr | Test Acc-2 | F1 | Non0 | Acc-3 | Acc-5 | MAE | Corr | Zero-F1 | Router [T,A,V] |
| --- | ---: | ---: | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | --- |
| `fde_tokenres_h160_a12_lr5e5_b8_d15` | 20262000 | 37 | ema | 0.732730 | 0.7785 | 0.3067 | 0.7852 | 0.8381 | 0.8408 | 0.8763 | 0.7505 | 0.5383 | 0.2845 | 0.8175 | 0.2742 | [0.466760, 0.277860, 0.255380] |
| `fde_tokenres_h160_a12_lr5e5_b8_d15_seed1` | 20262001 | 9 | ema | 0.732130 | 0.7763 | 0.3133 | 0.7879 | 0.8228 | 0.8264 | 0.8737 | 0.7593 | 0.5317 | 0.2971 | 0.7949 | 0.3540 | [0.421370, 0.300440, 0.278190] |
| `fde_tokenres_h160_a12_lr5e5_b8_d15_seed2` | 20262002 | 15 | ema | 0.713630 | 0.8004 | 0.3199 | 0.7784 | 0.8403 | 0.8424 | 0.8789 | 0.7527 | 0.5274 | 0.3077 | 0.7814 | 0.3220 | [0.461143, 0.316541, 0.222316] |
| `fde_tokenres_h160_a12_lr5e5_b8_d15_seed3` | 20262003 | 21 | ema | 0.735120 | 0.8070 | 0.3062 | 0.7913 | 0.8512 | 0.8523 | 0.8866 | 0.7484 | 0.5383 | 0.2956 | 0.7916 | 0.3193 | [0.457794, 0.282133, 0.260073] |

4-run mean/std：

| Metric | Mean | Std |
| --- | ---: | ---: |
| Mult_acc_2 | 0.8381 | 0.0117 |
| Non0_acc_2 | 0.8789 | 0.0056 |
| Mult_acc_5 | 0.5339 | 0.0053 |
| MAE | 0.2962 | 0.0095 |
| Corr | 0.7964 | 0.0152 |
| Zero_F1 | 0.3174 | 0.0328 |

## 对比 0617 CHSIMS Search 1

0617 CHSIMS Search 1 的最佳结果：

| Run | Test Acc-2 | Non0 | Acc-5 | MAE | Corr | Zero-F1 |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| `chsims_eng_h128_lr1e4_b8_d12` | 0.8249 | 0.9046 | 0.7155 | 0.2878 | 0.8137 | 0.6837 |
| `chsims_eng_h128_lr1e4_b12_d12` | 0.7987 | 0.8763 | 0.7002 | 0.2898 | 0.7925 | 0.6633 |
| `chsims_eng_h128_lr5e5_b8_d15` | 0.8118 | 0.8840 | 0.6980 | 0.3065 | 0.7993 | 0.6597 |
| `chsims_eng_h160_lr5e5_b8_d15` | 0.8074 | 0.8918 | 0.7133 | 0.3041 | 0.7980 | 0.6753 |

相对 0617 最强主结果 `chsims_eng_h128_lr1e4_b8_d12`：

| Metric | 0620 FDE mean | 0617 best | Delta |
| --- | ---: | ---: | ---: |
| Acc-2 | 0.8381 | 0.8249 | +0.0132 |
| Non0 | 0.8789 | 0.9046 | -0.0257 |
| Acc-5 | 0.5339 | 0.7155 | -0.1816 |
| MAE | 0.2962 | 0.2878 | +0.0084 |
| Corr | 0.7964 | 0.8137 | -0.0173 |
| Zero-F1 | 0.3174 | 0.6837 | -0.3663 |

相对 0617 同容量同 lr 的 `chsims_eng_h160_lr5e5_b8_d15`：

| Metric | 0620 FDE mean | 0617 h160 | Delta |
| --- | ---: | ---: | ---: |
| Acc-2 | 0.8381 | 0.8074 | +0.0307 |
| Non0 | 0.8789 | 0.8918 | -0.0129 |
| Acc-5 | 0.5339 | 0.7133 | -0.1794 |
| MAE | 0.2962 | 0.3041 | -0.0079 |
| Corr | 0.7964 | 0.7980 | -0.0016 |
| Zero-F1 | 0.3174 | 0.6753 | -0.3579 |

## 结论

- `fde_tokenres_h160_a12_lr5e5_b8_d15` 可以稳定迁移到 CHSIMS，4 个 run 均正常完成，并且均选择 EMA checkpoint。
- Acc-2 有明显提升，mean=0.8381，高于 0617 Search 1 最强主结果 0.8249；单 run seed3 达到 0.8512。
- Corr 单次最好为 0.8175，超过 0617 最强主结果的 0.8137；但 4-run mean Corr=0.7964，稳定性不足。
- 主要问题是分桶强度指标和中性类：Acc-5 mean=0.5339，Zero-F1 mean=0.3174，显著低于 0617 CHSIMS 基线。这说明该 MOSEI FDE 配置在 CHSIMS 上更偏向二分类符号判断，没有保住 SIMS 的强度分桶和 zero/neutral 辨识。

推荐：不要把 0620 FDE token-residual 作为 CHSIMS 主结果；它更适合作为“MOSEI 方法迁移到 CHSIMS 的对照/负结果”。如果继续做 CHSIMS，建议回到 0617 winning setting `h128/lr1e-4/b8/d12`，只把 FDE token-residual 作为小 alpha 消融加入，例如 `fd_token_alpha=0.04/0.08`，并保持 `desc_gate_mode=pre_align` 或加入 zero-weighted selection。
