diff --git a/docs/.vuepress/notes/en/guide.ts b/docs/.vuepress/notes/en/guide.ts
index 86a5eb9..d3654c3 100644
--- a/docs/.vuepress/notes/en/guide.ts
+++ b/docs/.vuepress/notes/en/guide.ts
@@ -24,6 +24,8 @@ export const Guide: ThemeNote = defineNoteConfig({
             items: [
                 'quickstart',
                 'tutorial',
+                'selector_delta_loss',
+                'selector_loss',
                 'selector_less',
                 'selector_nice',
                 'selector_offline_tsds',
diff --git a/docs/.vuepress/notes/zh/guide.ts b/docs/.vuepress/notes/zh/guide.ts
index 3bbd483..54e37ac 100644
--- a/docs/.vuepress/notes/zh/guide.ts
+++ b/docs/.vuepress/notes/zh/guide.ts
@@ -24,6 +24,8 @@ export const Guide: ThemeNote = defineNoteConfig({
             items: [
                 'quickstart',
                 'tutorial',
+                'selector_delta_loss',
+                'selector_loss',
                 'selector_less',
                 'selector_nice',
                 'selector_offline_tsds',
diff --git a/docs/en/notes/guide/mixer/odm.md b/docs/en/notes/guide/mixer/odm.md
index 8ba77ff..d8fea7d 100644
--- a/docs/en/notes/guide/mixer/odm.md
+++ b/docs/en/notes/guide/mixer/odm.md
@@ -1,7 +1,7 @@
 ---
 title: ODM Data Mixer
 createTime: 2025/01/27 10:00:00
-icon: material-symbols:casino
+icon: material-symbols:balance
 permalink: /en/guide/mixer/odm/
 ---
 
diff --git a/docs/en/notes/guide/selector/selector_delta_loss.md b/docs/en/notes/guide/selector/selector_delta_loss.md
new file mode 100644
index 0000000..e74d3d7
--- /dev/null
+++ b/docs/en/notes/guide/selector/selector_delta_loss.md
@@ -0,0 +1,183 @@
+---
+title: Delta Loss Selector
+createTime: 2025/12/27 00:44:11
+permalink: /en/guide/delta-loss/
+icon: carbon:select-window
+---
+# Delta Loss Selector Guide
+
+This document explains how to use the **Delta Loss Selector** in **DataFlex**. It tracks loss reduction relative to an initial baseline and samples from a sliding window over the ranked delta-loss list.
+
+---
+
+## 1. Method Overview
+
+**Delta Loss Selector** workflow:
+
+1. On the first selection, compute and cache **initial losses**, then return a random warmup batch.
+2. On later steps, compute current losses and define $\Delta l_i = l_i^{(init)} - l_i^{(current)}$.
+3. Sort samples by $\Delta l_i$ in descending order and compute a sliding-window position based on training progress.
+4. Assign high sampling probability inside the window and a small base probability outside.
+
+**Sliding window schedule:**
+
+Let update index be $t$ out of $T$, and window ratio be $s$:
+
+$$
+u = \sigma\left(\frac{t}{T}\right), \quad
+\text{start} = u \cdot (N - sN), \quad
+\text{end} = \text{start} + sN
+$$
+
+If $\Delta l_i < 0$, the window end is truncated to avoid prioritizing samples that got worse.
+
+
+## 2. Implementation Steps
+
+### Step 1: Environment Setup
+
+```bash
+git clone https://github.com/OpenDCAI/DataFlex.git
+cd DataFlex
+pip install -e .
+pip install llamafactory
+```
+
+---
+
+### Step 2: Delta Loss Selector Configuration
+
+**Configuration file path:**
+```
+DataFlex/src/dataflex/configs/components.yaml
+```
+
+**Example configuration:**
+```yaml
+delta_loss:
+  name: delta_loss
+  params:
+    cache_dir: ../dataflex_saves/delta_loss_output
+    window_size: 0.2
+```
+
+**Parameter Description:**
+* `cache_dir`: Cache directory; the first step stores initial losses here.
+* `window_size`: Sliding window ratio for focused sampling.
+
+---
+
+### Step 3: Dynamic Training Configuration
+
+**Configuration file path:**
+```
+DataFlex/examples/train_lora/selectors/delta_loss.yaml
+```
+
+**Example configuration:**
+```yaml
+### model
+model_name_or_path: meta-llama/Llama-3.1-8B
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+lora_rank: 16
+lora_alpha: 8
+
+### dataset
+dataset: alpaca_en_demo
+template: llama3
+cutoff_len: 4096
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 0
+seed: 42
+
+### output
+output_dir: ../dataflex_saves/Llama-3.1-8B/delta_loss
+logging_steps: 10
+save_steps: 100
+plot_loss: true
+save_only_model: false
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 1
+learning_rate: 1.0e-4
+num_train_epochs: 1.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### Dataflex args
+train_type: dynamic_select
+components_cfg_file: src/dataflex/configs/components.yaml
+component_name: delta_loss
+warmup_step: 10
+update_step: 10
+update_times: 2
+
+eval_dataset: alpaca_zh_demo
+```
+
+---
+
+### Step 4: Run Training
+
+```bash
+FORCE_TORCHRUN=1 DISABLE_VERSION_CHECK=1 dataflex-cli train examples/train_lora/selectors/delta_loss.yaml
+```
+
+---
+
+### Step 5: Model Merge and Export
+
+**Configuration file path:**
+
+```
+DataFlex/examples/merge_lora/llama3_lora_sft.yaml
+```
+
+**Example configuration:**
+
+```yaml
+model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+adapter_name_or_path: ../dataflex_saves/Llama-3.1-8B/delta_loss
+template: llama3
+trust_remote_code: true
+
+export_dir: ../dataflex_saves/Llama-3.1-8B_lora_sft
+export_size: 5
+export_device: cpu  # choices: [cpu, auto]
+export_legacy_format: false
+```
+
+**Parameter Description:**
+
+* `model_name_or_path`: Model name or path used for training.
+* `adapter_name_or_path`: Output path of the LoRA adapter.
+* `export_dir`: Directory for saving the merged result of the fine-tuned model and LoRA adapter.
+
+Execute the export command:
+
+```bash
+llamafactory-cli export llama3_lora_sft.yaml
+```
+
+The merged model will be saved in:
+
+```
+/dataflex_saves/Llama-3.1-8B_lora_sft
+```
+
+
+## 3. Model Evaluation
+
+It is recommended to use the [DataFlow](https://github.com/OpenDCAI/DataFlow) [Model QA Evaluation Pipeline](https://opendcai.github.io/DataFlow-Doc/zh/guide/2k5wjgls/) for systematic evaluation of the generated model.
+
diff --git a/docs/en/notes/guide/selector/selector_loss.md b/docs/en/notes/guide/selector/selector_loss.md
new file mode 100644
index 0000000..661cc83
--- /dev/null
+++ b/docs/en/notes/guide/selector/selector_loss.md
@@ -0,0 +1,188 @@
+---
+title: Loss Selector
+createTime: 2025/12/27 00:44:11
+permalink: /en/guide/loss/
+icon: carbon:select-window
+---
+# Loss Selector Guide
+
+This document explains how to use the **Loss Selector** in **DataFlex**. The selector computes per-sample training loss, splits the distribution into low/medium/high bands using quantiles, and then samples with higher weight on a chosen band.
+
+---
+
+## 1. Method Overview
+
+**Core idea of the Loss Selector:**
+
+1. During training, compute the training loss for each sample. In a multi-GPU setting, results from different processes are aligned to the full dataset using the sample index (`idx`).
+
+   * In the current implementation, `batch_size = 1`, so the loss returned by the model corresponds to a per-sample loss.
+2. On the main process, collect and deduplicate all valid sample losses, and partition samples into **low / medium / high** loss regions using quantile thresholds.
+3. Assign a base weight of 1 to all valid samples, and apply an amplified weight `focus_weight` to samples in the specified focus region (`focus`).
+4. Smooth the weight distribution using a temperature parameter and perform random sampling according to the resulting probability distribution; when the number of valid samples is insufficient to meet the sampling requirement, automatically switch to sampling with replacement.
+
+**Sampling probability:**
+
+Let the loss of sample ($i$) be ($l_i$), the segment weight be ($w_i$), and the temperature be ($T$):
+
+$$
+ p_i = \frac{(w_i + \epsilon)^{1/T}}{\sum_j (w_j + \epsilon)^{1/T}} 
+$$
+
+
+## 2. Implementation Steps
+
+### Step 1: Environment Setup
+
+```bash
+git clone https://github.com/OpenDCAI/DataFlex.git
+cd DataFlex
+pip install -e .
+pip install llamafactory
+```
+
+---
+
+### Step 2: Loss Selector Configuration
+
+**Configuration file path:**
+```
+DataFlex/src/dataflex/configs/components.yaml
+```
+
+**Example configuration:**
+```yaml
+loss:
+  name: loss
+  params:
+    cache_dir: ../dataflex_saves/loss_output
+    focus: "medium"        # low | medium | high
+    focus_weight: 5.0
+    quantiles: [0.33, 0.66]
+    replacement: false
+    temperature: 1.0
+```
+
+**Parameter Description:**
+* `cache_dir`: Cache directory for selection results (`step_{id}.json` per step).
+* `focus`: Target band to up-weight (`low` / `medium` / `high`, default `high`).
+* `focus_weight`: Weight multiplier for the focus band.
+* `quantiles`: Split points for low/medium/high, values in `[0, 1]`.
+* `replacement`: Sample with replacement or not; auto-switches to replacement if needed.
+* `temperature`: Distribution sharpness; `>1` smooths, `<1` sharpens.
+
+---
+
+### Step 3: Dynamic Training Configuration
+
+**Configuration file path:**
+```
+DataFlex/examples/train_lora/selectors/loss.yaml
+```
+
+**Example configuration:**
+```yaml
+### model
+model_name_or_path: meta-llama/Llama-3.1-8B
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+lora_rank: 16
+lora_alpha: 8
+
+### dataset
+dataset: alpaca_en_demo
+template: llama3
+cutoff_len: 4096
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 0
+seed: 42
+
+### output
+output_dir: ../dataflex_saves/Llama-3.1-8B/loss
+logging_steps: 10
+save_steps: 100
+plot_loss: true
+save_only_model: false
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 1
+learning_rate: 1.0e-4
+num_train_epochs: 1.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### Dataflex args
+train_type: dynamic_select
+components_cfg_file: src/dataflex/configs/components.yaml
+component_name: loss
+warmup_step: 10
+update_step: 10
+update_times: 2
+
+eval_dataset: alpaca_zh_demo
+```
+
+---
+
+### Step 4: Run Training
+
+```bash
+FORCE_TORCHRUN=1 DISABLE_VERSION_CHECK=1 dataflex-cli train examples/train_lora/selectors/loss.yaml
+```
+
+
+### Step 5: Model Merge and Export
+
+**Configuration file path:**
+
+```
+DataFlex/examples/merge_lora/llama3_lora_sft.yaml
+```
+
+**Example configuration:**
+
+```yaml
+model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+adapter_name_or_path: ../dataflex_saves/Llama-3.1-8B/loss
+template: llama3
+trust_remote_code: true
+
+export_dir: ../dataflex_saves/Llama-3.1-8B_lora_sft
+export_size: 5
+export_device: cpu  # choices: [cpu, auto]
+export_legacy_format: false
+```
+
+**Parameter Description:**
+
+* `model_name_or_path`: Model name or path used for training.
+* `adapter_name_or_path`: Output path of the LoRA adapter.
+* `export_dir`: Directory for saving the merged result of the fine-tuned model and LoRA adapter.
+
+Execute the export command:
+
+```bash
+llamafactory-cli export llama3_lora_sft.yaml
+```
+
+The merged model will be saved in:
+
+```
+/dataflex_saves/Llama-3.1-8B_lora_sft
+```
+
+
+## 3. Model Evaluation
+
+It is recommended to use the [DataFlow](https://github.com/OpenDCAI/DataFlow) [Model QA Evaluation Pipeline](https://opendcai.github.io/DataFlow-Doc/zh/guide/2k5wjgls/) for systematic evaluation of the generated model.
+
diff --git a/docs/en/notes/guide/selector/selector_nice.md b/docs/en/notes/guide/selector/selector_nice.md
index 7249f80..5c77b0f 100644
--- a/docs/en/notes/guide/selector/selector_nice.md
+++ b/docs/en/notes/guide/selector/selector_nice.md
@@ -2,7 +2,7 @@
 title: NICE Data Selector
 createTime: 2025/12/17 12:00:08
 permalink: /en/guide/nice/
-icon: carbon:select-02
+icon: carbon:select-window
 ---
 # NICE Selector Usage Guide
 
diff --git a/docs/en/notes/guide/selector/selector_offline_near.md b/docs/en/notes/guide/selector/selector_offline_near.md
index 8173cd8..2e967c7 100644
--- a/docs/en/notes/guide/selector/selector_offline_near.md
+++ b/docs/en/notes/guide/selector/selector_offline_near.md
@@ -2,7 +2,7 @@
 title: Offline-Near-Selector
 createTime: 2025/11/27 16:02:41
 permalink: /en/guide/7k0w3d92/
-icon: flowbite:fish-alt-outline
+icon: carbon:select-02
 ---
 # Offline NEAR Selector 
 
diff --git a/docs/en/notes/guide/selector/selector_offline_tsds.md b/docs/en/notes/guide/selector/selector_offline_tsds.md
index 3e3af90..137ced9 100644
--- a/docs/en/notes/guide/selector/selector_offline_tsds.md
+++ b/docs/en/notes/guide/selector/selector_offline_tsds.md
@@ -2,7 +2,7 @@
 title: Offline-Tsds-Selector
 createTime: 2025/11/01 21:36:21
 permalink: /en/guide/im5q9cd2/
-icon: tdesign:cat
+icon: carbon:select-02
 ---
 
 
diff --git a/docs/en/notes/guide/selector/selector_zeroth.md b/docs/en/notes/guide/selector/selector_zeroth.md
index 101033c..5d588d3 100644
--- a/docs/en/notes/guide/selector/selector_zeroth.md
+++ b/docs/en/notes/guide/selector/selector_zeroth.md
@@ -2,7 +2,7 @@
 title: Zero Order Data Selection
 createTime: 2025/11/03 22:58:52
 permalink: /en/guide/0rfxa64a/
-icon: tabler:chart-dots-3
+icon: carbon:select-window
 ---
 # Introduction to the Zeroth Selector
 This document explains how to use the **Zeroth Selector** in the **DataFlex** framework to achieve dynamic selection of training data, thereby enhancing the performance of supervised fine-tuning (SFT). This method is an original selection approach that utilizes co-directional perturbations on the model for differential estimation, thereby obtaining the model's zeroth-order gradient to calculate the effective score of the data.
diff --git a/docs/zh/notes/guide/mixer/odm.md b/docs/zh/notes/guide/mixer/odm.md
index e06d41d..bd8ccd4 100644
--- a/docs/zh/notes/guide/mixer/odm.md
+++ b/docs/zh/notes/guide/mixer/odm.md
@@ -1,7 +1,7 @@
 ---
 title: ODM 数据混合器
 createTime: 2025/01/27 10:00:00
-icon: material-symbols:casino
+icon: material-symbols:balance
 permalink: /zh/guide/mixer/odm/
 ---
 
diff --git a/docs/zh/notes/guide/selector/selector_delta_loss.md b/docs/zh/notes/guide/selector/selector_delta_loss.md
new file mode 100644
index 0000000..d89333d
--- /dev/null
+++ b/docs/zh/notes/guide/selector/selector_delta_loss.md
@@ -0,0 +1,181 @@
+---
+title: Delta Loss 数据选择器
+createTime: 2025/12/27 00:44:11
+permalink: /zh/guide/delta-loss/
+icon: carbon:select-window
+---
+# Delta Loss Selector 使用介绍
+
+本文档介绍如何在 **DataFlex** 框架中使用 **Delta Loss Selector**。该方法以“初始损失 - 当前损失”作为样本有效性的信号，结合滑动窗口策略进行动态采样，从而在训练过程中逐步移动关注区间。
+
+---
+
+## 1. 方法概述
+
+**Delta Loss Selector** 的核心流程：
+
+1. 首次选择时计算并缓存 **初始损失**，并随机采样作为 warmup。
+2. 后续步骤计算当前损失，得到 $\Delta l_i = l_i^{(init)} - l_i^{(current)}$。
+3. 对 $\Delta l_i$ 降序排序，并按训练进度计算滑动窗口位置。
+4. 对窗口内样本赋较高采样概率，窗口外样本保留较低基准概率。
+
+**滑动窗口位置：**
+
+设当前更新次数为 $t$，总更新次数为 $T$，窗口大小为 $s$（比例），
+窗口起点通过 Sigmoid 调度获得：
+
+$$
+u = \sigma\left(\frac{t}{T}\right), \quad
+\text{start} = u \cdot (N - sN), \quad
+\text{end} = \text{start} + sN
+$$
+
+当 $\Delta l_i < 0$ 时，窗口右端会被截断，避免选择损失变差的样本。
+
+
+## 2. 实现步骤
+
+### 步骤一：环境安装
+
+```bash
+git clone https://github.com/OpenDCAI/DataFlex.git
+cd DataFlex
+pip install -e .
+pip install llamafactory
+```
+
+---
+
+### 步骤二：Delta Loss Selector 参数配置
+
+**配置文件路径：**
+```
+DataFlex/src/dataflex/configs/components.yaml
+```
+
+**示例配置：**
+```yaml
+delta_loss:
+  name: delta_loss
+  params:
+    cache_dir: ../dataflex_saves/delta_loss_output
+    window_size: 0.2
+```
+
+**参数说明：**
+* `cache_dir`: 选择结果缓存路径，首次选择会保存 initial loss。
+* `window_size`: 滑动窗口大小（比例），决定每轮重点采样的区间宽度。
+
+---
+
+### 步骤三：动态训练配置
+
+**配置文件路径：**
+```
+DataFlex/examples/train_lora/selectors/delta_loss.yaml
+```
+
+**示例配置：**
+```yaml
+### model
+model_name_or_path: meta-llama/Llama-3.1-8B
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+lora_rank: 16
+lora_alpha: 8
+
+### dataset
+dataset: alpaca_en_demo
+template: llama3
+cutoff_len: 4096
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 0
+seed: 42
+
+### output
+output_dir: ../dataflex_saves/Llama-3.1-8B/delta_loss
+logging_steps: 10
+save_steps: 100
+plot_loss: true
+save_only_model: false
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 1
+learning_rate: 1.0e-4
+num_train_epochs: 1.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### Dataflex args
+train_type: dynamic_select
+components_cfg_file: src/dataflex/configs/components.yaml
+component_name: delta_loss
+warmup_step: 10
+update_step: 10
+update_times: 2
+
+eval_dataset: alpaca_zh_demo
+```
+
+---
+
+### 步骤四：运行训练
+
+```bash
+FORCE_TORCHRUN=1 DISABLE_VERSION_CHECK=1 dataflex-cli train examples/train_lora/selectors/delta_loss.yaml
+```
+
+---
+
+### 步骤五：模型合并与导出
+
+**配置文件路径：**
+
+```
+DataFlex/examples/merge_lora/llama3_lora_sft.yaml
+```
+
+**示例配置：**
+
+```yaml
+model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+adapter_name_or_path: ../dataflex_saves/Llama-3.1-8B/delta_loss
+template: llama3
+trust_remote_code: true
+
+export_dir: ../dataflex_saves/Llama-3.1-8B_lora_sft
+export_size: 5
+export_device: cpu  # choices: [cpu, auto]
+export_legacy_format: false
+```
+**参数说明：**
+* `model_name_or_path`: 训练模型的名称或路径。
+* `adapter_name_or_path`: LoRA适配器输出路径。
+* `export_dir`: 监督微调后的模型，训练模型与LoRA适配器的合并结果。
+
+执行合并导出命令：
+
+```bash
+llamafactory-cli export llama3_lora_sft.yaml
+```
+
+合并后的模型将保存在如下文件夹：
+
+```
+/dataflex_saves/Llama-3.1-8B_lora_sft
+```
+
+## 3. 模型评估
+
+推荐使用[DataFlow](https://github.com/OpenDCAI/DataFlow)的[模型QA能力评估流水线](https://opendcai.github.io/DataFlow-Doc/zh/guide/2k5wjgls/)对生成后的模型进行系统性评估。
+
diff --git a/docs/zh/notes/guide/selector/selector_loss.md b/docs/zh/notes/guide/selector/selector_loss.md
new file mode 100644
index 0000000..205a5a8
--- /dev/null
+++ b/docs/zh/notes/guide/selector/selector_loss.md
@@ -0,0 +1,187 @@
+---
+title: Loss 数据选择器
+createTime: 2025/12/27 00:44:11
+permalink: /zh/guide/loss/
+icon: carbon:select-window
+---
+# Loss Selector 使用介绍
+
+本文档介绍如何在 **DataFlex** 框架中使用 **Loss Selector**，基于样本损失分布进行动态数据选择。该方法会按分位数将样本划分为低/中/高损失区间，并对指定区间加权采样，以便在训练中更聚焦特定难度的样本。
+
+---
+
+## 1. 方法概述
+
+
+**Loss Selector** 的核心思想是：
+1. 在训练过程中计算样本对应的训练损失，并在多卡环境下通过样本索引（`idx`）将各进程结果对齐到完整数据集。
+
+   * 当前实现使用 `batch_size=1`，因此模型返回的 loss 等价于逐样本损失。
+2. 在主进程上收集并去重所有样本的有效损失值，使用分位数阈值将样本划分为 **low / medium / high** 三个损失区间。
+3. 为所有有效样本赋予基础权重 1，并对指定关注区间（`focus`）内的样本施加放大权重 `focus_weight`。
+4. 通过温度参数对权重分布进行平滑，并依据得到的概率分布进行随机采样；当有效样本数量不足以满足采样需求时，自动切换为放回采样。
+
+**采样概率：**
+
+设样本损失为 $l_i$，分段权重为 $w_i$，温度为 $T$： 
+
+$$
+ p_i = \frac{(w_i + \epsilon)^{1/T}}{\sum_j (w_j + \epsilon)^{1/T}} 
+$$
+
+
+## 2. 实现步骤
+
+### 步骤一：环境安装
+
+```bash
+git clone https://github.com/OpenDCAI/DataFlex.git
+cd DataFlex
+pip install -e .
+pip install llamafactory
+```
+
+---
+
+### 步骤二：Loss Selector 参数配置
+
+**配置文件路径：**
+```
+DataFlex/src/dataflex/configs/components.yaml
+```
+
+**示例配置：**
+```yaml
+loss:
+  name: loss
+  params:
+    cache_dir: ../dataflex_saves/loss_output
+    focus: "medium"        # low | medium | high
+    focus_weight: 5.0
+    quantiles: [0.33, 0.66]
+    replacement: false
+    temperature: 1.0
+```
+
+**参数说明：**
+* `cache_dir`: 选择结果缓存路径（每个 step 会写入 `step_{id}.json`）。
+* `focus`: 关注区间，可选 `low` / `medium` / `high`（默认 `high`）。
+* `focus_weight`: 关注区间的权重倍数，越大越偏向该区间。
+* `quantiles`: 低/中/高损失的分位数切分点，取值在 `[0, 1]`。
+* `replacement`: 是否放回采样；若请求数大于有效样本量，会自动切换为放回采样。
+* `temperature`: 温度系数，`>1` 更平滑，`<1` 更尖锐。
+
+---
+
+### 步骤三：动态训练配置
+
+**配置文件路径：**
+```
+DataFlex/examples/train_lora/selectors/loss.yaml
+```
+
+**示例配置：**
+```yaml
+### model
+model_name_or_path: meta-llama/Llama-3.1-8B
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+lora_rank: 16
+lora_alpha: 8
+
+### dataset
+dataset: alpaca_en_demo
+template: llama3
+cutoff_len: 4096
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_num_workers: 0
+seed: 42
+
+### output
+output_dir: ../dataflex_saves/Llama-3.1-8B/loss
+logging_steps: 10
+save_steps: 100
+plot_loss: true
+save_only_model: false
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 1
+learning_rate: 1.0e-4
+num_train_epochs: 1.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### Dataflex args
+train_type: dynamic_select
+components_cfg_file: src/dataflex/configs/components.yaml
+component_name: loss
+warmup_step: 10
+update_step: 10
+update_times: 2
+
+eval_dataset: alpaca_zh_demo
+```
+
+---
+
+### 步骤四：运行训练
+
+```bash
+FORCE_TORCHRUN=1 DISABLE_VERSION_CHECK=1 dataflex-cli train examples/train_lora/selectors/loss.yaml
+```
+
+---
+
+
+### 步骤五：模型合并与导出
+
+**配置文件路径：**
+
+```
+DataFlex/examples/merge_lora/llama3_lora_sft.yaml
+```
+
+**示例配置：**
+
+```yaml
+model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+adapter_name_or_path: ../dataflex_saves/Llama-3.1-8B/less
+template: llama3
+trust_remote_code: true
+
+export_dir: ../dataflex_saves/Llama-3.1-8B_lora_sft
+export_size: 5
+export_device: cpu  # choices: [cpu, auto]
+export_legacy_format: false
+```
+**参数说明：**
+* `model_name_or_path`: 训练模型的名称或路径。
+* `adapter_name_or_path`: LoRA适配器输出路径。
+* `export_dir`: 监督微调后的模型，训练模型与LoRA适配器的合并结果。
+
+执行合并导出命令：
+
+```bash
+llamafactory-cli export llama3_lora_sft.yaml
+```
+
+合并后的模型将保存在如下文件夹：
+
+```
+/dataflex_saves/Llama-3.1-8B_lora_sft
+```
+
+## 3. 模型评估
+
+推荐使用[DataFlow](https://github.com/OpenDCAI/DataFlow)的[模型QA能力评估流水线](https://opendcai.github.io/DataFlow-Doc/zh/guide/2k5wjgls/)对生成后的模型进行系统性评估。
+
diff --git a/docs/zh/notes/guide/selector/selector_nice.md b/docs/zh/notes/guide/selector/selector_nice.md
index d175a96..51c8ea0 100644
--- a/docs/zh/notes/guide/selector/selector_nice.md
+++ b/docs/zh/notes/guide/selector/selector_nice.md
@@ -2,7 +2,7 @@
 title: Nice 数据选择器
 createTime: 2025/12/17 12:00:08
 permalink: /zh/guide/nice/
-icon: carbon:select-02
+icon: carbon:select-window
 ---
 # NICE Selector 使用介绍
 
diff --git a/docs/zh/notes/guide/selector/selector_offline_near.md b/docs/zh/notes/guide/selector/selector_offline_near.md
index 8fb89f9..6de560e 100644
--- a/docs/zh/notes/guide/selector/selector_offline_near.md
+++ b/docs/zh/notes/guide/selector/selector_offline_near.md
@@ -2,7 +2,7 @@
 title: Offline-Near数据选择器
 createTime: 2025/11/26 23:42:41
 permalink: /zh/guide/acgesu99/
-icon: flowbite:fish-alt-outline
+icon: carbon:select-02
 ---
 # Offline NEAR Selector 使用介绍
 
diff --git a/docs/zh/notes/guide/selector/selector_offline_tsds.md b/docs/zh/notes/guide/selector/selector_offline_tsds.md
index b255994..9bdaa2e 100644
--- a/docs/zh/notes/guide/selector/selector_offline_tsds.md
+++ b/docs/zh/notes/guide/selector/selector_offline_tsds.md
@@ -2,7 +2,7 @@
 title: Offline-Tsds 数据选择器
 createTime: 2025/11/01 21:35:45
 permalink: /zh/guide/vkqfowej/
-icon: tdesign:cat
+icon: carbon:select-02
 
 ---
 
diff --git a/docs/zh/notes/guide/selector/selector_zeroth.md b/docs/zh/notes/guide/selector/selector_zeroth.md
index 7aef476..ba59137 100644
--- a/docs/zh/notes/guide/selector/selector_zeroth.md
+++ b/docs/zh/notes/guide/selector/selector_zeroth.md
@@ -2,7 +2,7 @@
 title: 零阶优化数据选择器
 createTime: 2025/11/03 22:58:52
 permalink: /zh/guide/wl9tf7o2/
-icon: tabler:chart-dots-3
+icon: carbon:select-window
 ---
 # 零阶选择器介绍
 本文档介绍如何在 **DataFlex** 框架中使用 **Zeroth Selector** 实现训练数据的动态选择，从而提升监督微调（SFT）效果。该方法为原创选择方法，利用对模型进行同向扰动进行差分估计，进而得到模型的零阶梯度来计算数据有效分数。