### Install Requirements

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/experiments/ernie_pretrain/README.md

Installs necessary Python packages for the pre-training environment. Use --force-reinstall to ensure a clean installation.

```bash
mpirun python -m pip install -r requirements.txt --force-reinstall
```

--------------------------------

### Start VisualDL for Training Visualization

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/PaddleOCR-VL/README.md

Starts the VisualDL service to visualize training logs. Ensure the specified port is available.

```shell
visualdl --logdir ./PaddleOCR-VL-SFT-Bengali/visualdl_logs/ --port 8084
```

--------------------------------

### Start Model Training

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/tutorials/how_to_train_a_visual_grounding_model.md

Initiate model training with a single command. Ensure the configuration file (e.g., qwen25vl_grounding_lora.yaml) is correctly specified.

```shell
CUDA_VISIBLE_DEVICES=0 paddleformers-cli train qwen25vl_grounding_lora.yaml
```

--------------------------------

### Example Prompts for Model Training

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/tutorials/how_to_train_an_emoji_model.md

These are example prompts used to demonstrate the model's initial response before fine-tuning for emoji integration.

```python
prompt="写一篇关于La Taqueria餐厅的评论。"
prompt="比较美国和加拿大的学生债务危机。"
```

--------------------------------

### Install PaddleFormers

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/DeepSeek-V3/SFT-Practice.md

Clone the PaddleFormers repository and install the necessary dependencies and the package itself.

```bash
git clone https://github.com/PaddlePaddle/PaddleFormers.git
pip install -r requirements.txt
pip install -e .
```

--------------------------------

### Launch Pre-Training on 2016 GPUs

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/experiments/ernie_pretrain/README.md

Starts the pre-training process using a script designed for 2016 GPUs. Ensure master IP and port are configured correctly in the script.

```bash
mpirun bash scripts/train_2016_gpus.sh
```

--------------------------------

### Install a Module with PaddleHub

Source: https://github.com/paddlepaddle/paddleformers/wiki/PaddleHub命令行工具

Installs a Module to the local directory. By default, it is installed in `${HUB_HOME}/.paddlehub/modules`.

```shell
hub install <module_name>
```

--------------------------------

### Verify PaddleFormers CLI Installation

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/docs/en/cli_usage.md

Check if the PaddleFormers CLI is installed correctly by running the help command.

```bash
paddleformers-cli help
```

--------------------------------

### Launch SFT-LoRA Training

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/ERNIE-4.5-VL/README.md

Starts the SFT-LoRA training process with a specific configuration file. This method is less resource-intensive, requiring 4 x 80GB GPUs.

```bash
# SFT-LoRA训练需要4 * 80G GPU
CUDA_VISIBLE_DEVICES=0,1,2,3 paddleformers-cli train examples/best_practices/Ernie4.5VL/ernie45vl_8k_lora_config.yaml
```

--------------------------------

### Launch Pre-Training on 96 GPUs

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/experiments/ernie_pretrain/README.md

Starts the pre-training process using a script designed for 96 GPUs. Ensure master IP and port are configured correctly in the script.

```bash
mpirun bash scripts/train_96_gpus.sh
```

--------------------------------

### Launch TensorBoard for Visualization

Source: https://github.com/paddlepaddle/paddleformers/wiki/PaddleHub文本分类迁移教程

Starts TensorBoard to visualize the fine-tuning progress, including loss and accuracy metrics. Replace `${HOST_IP}` and `${PORT}` with your machine's IP and a desired port.

```shell
$ tensorboard --logdir ./cv_finetune_turtorial_demo/visualization --host ${HOST_IP} --port ${PORT}
```

--------------------------------

### Install PaddleFormers CLI

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/docs/en/cli_usage.md

Install the PaddleFormers package in editable mode from the root directory.

```bash
python -m pip install -e .
```

--------------------------------

### Launch VisualDL for Training Monitoring

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/PaddleOCR-VL-1.5/region_ocr.md

Starts VisualDL to monitor the training process. The `--logdir` should point to your VisualDL log directory, and `--port` should be set to an available port.

```bash
visualdl --logdir ./PaddleOCR-VL-1.5-SFT-RegionOCR/visualdl_logs/ --port 8084
```

--------------------------------

### Start Multi-Node/Multi-GPU Training

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/ERNIE-4.5/README.md

Launch pre-training for the 300B ERNIE-4.5 model on multiple nodes and GPUs. Set environment variables for distributed training.

```shell
NNODES={num_nodes} \
MASTER_ADDR={your_master_addr} \
MASTER_PORT={your_master_port} \
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
paddleformers-cli train examples/config/pt/eb45_pretrain/300b_96gpus.yaml
```

--------------------------------

### Iterate Through Training Examples

Source: https://github.com/paddlepaddle/paddleformers/wiki/PaddleHub文本分类迁移教程

Demonstrates how to iterate through the training examples of the ChnSentiCorp dataset to inspect text and labels.

```python
ds = hub.dataset.ChnSentiCorp()
for e in ds.get_train_examples():
    print(e.text_a, e.label)
```

--------------------------------

### Start Single-Node 8-GPU Training

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/ERNIE-4.5/README.md

Launch pre-training for the 21B ERNIE-4.5 model on a single machine with 8 GPUs using the PaddleFormers CLI.

```shell
paddleformers-cli train examples/config/pt/eb45_pretrain/21b_8_gpus.yaml
```

--------------------------------

### Train Model with LoRA (Pre-training)

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/README.md

Start LoRA pre-training using a specified configuration file. Supports single-card and multi-card training with online data streams.

```bash
# Single-card, online data stream
paddleformers-cli train ./config/pt/lora.yaml

# Multi-card, online data stream
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 paddleformers-cli train ./config/pt/lora_tp_pp.yaml
```

--------------------------------

### Load and Process Image with Qwen2.5-VL-3B-Instruct

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/docs/en/image_processors.md

Loads an Image Processor for the Qwen2.5-VL-3B-Instruct model and processes a sample image, returning the output as Paddle Tensors. Ensure PIL and requests are installed.

```python
from paddleformers.transformers import AutoImageProcessor
from PIL import Image
import requests

image_processor = AutoImageProcessor.from_pretrained("Qwen/Qwen2.5-VL-3B-Instruct")

url = "https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_images/example1.jpg"
image = Image.open(requests.get(url, stream=True).raw).convert("RGB")

processed_image = image_processor(image, return_tensors="pd")   # return Paddle Tensor
```

--------------------------------

### Function Calling Dataset Example

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/tutorials/how_to_train_a_function_call_model.md

This JSONL example demonstrates the structure of a single data entry for training a Function Calling model. It includes user messages, assistant responses with tool calls, and the available tools.

```json
{"messages": [{"role": "user", "content": "I need to generate an invoice for a customer named John Doe. He bought 2 apples for $1 each and 3 oranges for $2 each."}, {"role": "assistant", "content": "<think>\n</think>", "tool_calls": [{"type": "function", "function": {"name": "generate_invoice", "arguments": {"customer_name": "John Doe", "items": [{"name": "apple", "quantity": 2, "price": 1}, {"name": "orange", "quantity": 3, "price": 2}]}}}]}], "tools": [{"type": "function", "function": {"name": "generate_invoice", "description": "Generate an invoice for a customer", "parameters": {"type": "object", "properties": {"customer_name": {"type": "string", "description": "Name of the customer"}, "items": {"type": "array", "items": {"type": "object", "properties": {"name": {"type": "string", "description": "Name of the item"}, "quantity": {"type": "integer", "description": "Quantity of the item"}, "price": {"type": "number", "description": "Price of the item"}}, "required": ["name", "quantity", "price"]}, "description": "List of items in the invoice"}}, "required": ["customer_name", "items"]}}}]}
```

--------------------------------

### NLP Custom Dataset Preparation Example

Source: https://github.com/paddlepaddle/paddleformers/wiki/PaddleHub适配自定义数据完成finetune

Example of a custom dataset file format for NLP tasks, with text and labels separated by tabs. The first line specifies the column headers.

```text
text_a	label
15.4寸笔记本的键盘确实爽，基本跟台式机差不多了，蛮喜欢数字小键盘，输数字特方便，样子也很美观，做工也相当不错	1
房间太小。其他的都一般。。。。。。。。。	0
1.接电源没有几分钟,电源适配器热的不行. 2.摄像头用不起来. 3.机盖的钢琴漆，手不能摸，一摸一个印. 4.硬盘分区不好办.	0
```

--------------------------------

### Start Fine-tuning and Evaluation

Source: https://github.com/paddlepaddle/paddleformers/wiki/PaddleHub文本分类迁移教程

Initiates the fine-tuning process using the `finetune_and_eval` method. This automatically handles validation, saves the best model, and sets up TensorBoard visualization.

```python
cls_task.finetune_and_eval()
```

--------------------------------

### Create Baseline JSON File

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/scripts/xpu_ci/README.md

Example JSON structure for a baseline file. Keys represent step numbers and values represent the corresponding loss values.

```json
{
    "1": 7.12345678,
    "2": 7.08901234,
    ...
}
```

--------------------------------

### Evaluation Script Setup

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/tutorials/how_to_train_a_visual_grounding_model.md

Python script for evaluating model performance by comparing predictions with ground truth. Includes necessary imports and data structures.

```python
import json
import math
import os
import re

from PIL import Image, ImageDraw, ImageFont


MODEL_RESULT = {
    "image_path": "images/000000299887.jpg",
    "ground_truth": {
        "ref": [
            "motorcycle",
            "person",
            "person",
            "truck"
        ],
        "bbox": [
            [5, 129, 322, 471],
            [212, 105, 447, 475],
            [373, 121, 522, 475],
            [0, 178, 48, 213]
        ]
    },
    "prediction": "motorcycle(0,132),(379,475), person(212,106),(442,475), person(362,118),(522,475), truck(0,189),(48,213)"
}
ROOT_DIR = "data/coco_grounding"
```

--------------------------------

### ChatML Format Example

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/docs/en/datasets_format.md

Illustrates the ChatML format for conversational data, including roles like 'tool' and 'assistant', and tool definitions for function calls.

```json
{
            "role": "tool", 
            "content": [{"joke": "Why don't scientists trust atoms? Because they make up everything!"}]},
            {"role": "assistant", "content": "Sure, here's a joke for you: \"Why don't scientists trust atoms? Because they make up everything!\" I hope that brings a smile to your face."
        }

```

```json
        ],
        "tools": [
            {"type": "function", "function": {"name": "get_random_joke", "description": "Get a random joke", "parameters": {"type": "object", "properties": {}, "required": []}}},
            {"type": "function", "function": {"name": "generate_random_number", "description": "Generate a random number within a specified range", "parameters": {"type": "object", "properties": {"min": {"type": "number", "description": "The minimum value of the range"}, "max": {"type": "number", "description": "The maximum value of the range"}}, "required": ["min", "max"]}}}
        ]
    }

```

--------------------------------

### Start LoRA SFT Training

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/PaddleOCR-VL/README.md

Launches LoRA-based SFT training for PaddleOCR-VL. Similar to full parameter training, `pre_alloc_memory` helps manage GPU memory.

```shell
CUDA_VISIBLE_DEVICES=0 \
paddleformers-cli train examples/best_practices/PaddleOCR-VL/paddleocr-vl_lora_16k_config.yaml \
                        model_name_or_path=PaddlePaddle/PaddleOCR-VL \
                        train_dataset_path=./ocr_vl_sft-train_Bengali.jsonl \
                        eval_dataset_path=./ocr_vl_sft-test_Bengali.jsonl\
                        pre_alloc_memory=16
```

--------------------------------

### Launch Training with PaddleFormers CLI

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/DeepSeek-V3/README.md

Initiate the training process using the PaddleFormers CLI with a specified configuration file. This command is suitable for smaller-scale or example training runs.

```shell
# PaddleFormers/examples/best_practices/DeepseekV3_Pretrain/
# 可参考 run.sh、train_gpu.sh
paddleformers-cli train ./config/pretrain_argument.yaml
```

--------------------------------

### Launch Training Script with mpirun

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/DeepSeek-V3/SFT-Practice.md

Execute the training script for a 4K context length using mpirun for multi-machine startup. Ensure the corresponding script file is available.

```bash
# 推荐使用mpirun进行多机启动，以4k为例
mpirun bash run_dsv3_4k.sh
```

--------------------------------

### Install PaddlePaddle GPU Version

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/DeepSeek-V3/SFT-Practice.md

Install the PaddlePaddle GPU version, specifying the CUDA version. Refer to the official documentation for other installation options.

```bash
# cuda12.9
python -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu129/
#更多安装版本参考 https://www.paddlepaddle.org.cn/
```

--------------------------------

### Install PaddlePaddle with GPU Support

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/ERNIE-4.5-VL/README.md

Installs the PaddlePaddle deep learning framework with GPU support for CUDA 12.6. Refer to the official documentation for other installation versions.

```bash
# cuda12.6
python -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
```

--------------------------------

### List Installed Modules with PaddleHub

Source: https://github.com/paddlepaddle/paddleformers/wiki/PaddleHub命令行工具

Lists all Modules that have been installed locally.

```shell
hub list
```

--------------------------------

### Install PaddlePaddle GPU Version

Source: https://github.com/paddlepaddle/paddleformers/wiki/PaddleHub-FAQ

Install the GPU version of PaddlePaddle if your machine supports it.

```shell
$ pip install paddlepaddle-gpu
```

--------------------------------

### Launch SFT-FULL Training

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/ERNIE-4.5-VL/README.md

Initiates the SFT-FULL training process using a specified configuration file. This command requires 8 x 80GB GPUs and uses the paddleformers-cli tool.

```bash
# SFT-FULL训练需要8 * 80G GPU
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 paddleformers-cli train examples/best_practices/Ernie4.5VL/ernie45vl_8k_config.yaml
```

--------------------------------

### Install PaddlePaddle CPU Version

Source: https://github.com/paddlepaddle/paddleformers/wiki/PaddleHub-FAQ

Install the CPU version of PaddlePaddle if your machine does not support GPU.

```shell
$ pip install paddlepaddle
```

--------------------------------

### Train Model with Full Parameters (Pre-training)

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/README.md

Initiate pre-training with full parameters using a specified configuration file. Supports single-card and multi-card training with online or offline data streams.

```bash
# Single-card, online data stream
paddleformers-cli train ./config/pt/full.yaml

# Multi-card, online data stream
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 paddleformers-cli train ./config/pt/full_tp_pp.yaml

# Single-card, offline data stream
paddleformers-cli train ./config/pt/full_offline_data.yaml
```

--------------------------------

### Example Training Command Output

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/scripts/xpu_ci/README.md

Illustrates the detailed output when a training command is executed, including command, working directory, config, baseline, tolerance, timeout, and log file information.

```text
================================================================================
EXECUTING TRAINING COMMAND
================================================================================
Command: paddleformers-cli train tests/xpu_ci/config/ernie_21b_sft.yaml
Working Directory: /paddle/sjx_cuda12.6_py310/formers_test/PaddleFormers
Config File: tests/xpu_ci/config/ernie_21b_sft.yaml
Baseline File: tests/xpu_ci/base_value/ernie_21b_sft_loss.json
Tolerance: 1e-06
Timeout: 3600s
Log File: /tmp/pytest-xxx/test_ernie_21b_sft_training0/test_ernie_21b_sft_training.log
================================================================================

Command finished with return code: 0
✓ Command succeeded. Log saved to: /tmp/pytest-xxx/test_ernie_21b_sft_training0/test_ernie_21b_sft_training.log

✓ Loss validation PASSED - all values within tolerance
```

--------------------------------

### Download Pre-training Demo Dataset

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/docs/en/datasets_format.md

Use wget to download the pre-training demo dataset. This dataset can be used directly for testing.

```shell
wget https://paddleformers.bj.bcebos.com/datasets/pt_data.tar.gz
mkdir -p data/pt && tar -xf pt_data.tar.gz -C data/pt/
```

--------------------------------

### Multi-node Training Method 2 (mpirun)

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/docs/en/cli_usage.md

Set up multi-node training using mpirun by first creating a shell script containing the training command with necessary environment variables, then executing it via mpirun.

```bash
# First, write a script, such as scripts/train_96_gpus.sh, with the following content:
NNODES={num_nodes} MASTER_ADDR={your_master_addr} MASTER_PORT={your_master_port} RANK={rank} CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 paddleformers-cli train examples/config/sft_full.yaml
```

```bash
# Then:
mpirun bash scripts/train_96_gpus.sh
```

--------------------------------

### Install PaddleFormers

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/ERNIE-4.5-VL/README.md

Clones the PaddleFormers repository and installs it in editable mode. This allows for direct modification and use of the library's components.

```bash
git clone https://github.com/PaddlePaddle/PaddleFormers.git
cd PaddleFormers
pip install -e .
```

--------------------------------

### Train Reasoning Model via Command Line

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/tutorials/how_to_train_a_reasoning_model.md

Initiate model training by executing this command. Ensure the YAML configuration file is correctly specified.

```shell
paddleformers-cli train thinking.yaml
```

--------------------------------

### Install PaddleX for OCR

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/PaddleOCR-VL-1.5/table_ocr.md

Install the PaddleX library with OCR support using pip. This is a prerequisite for using the OTSL to HTML conversion script.

```bash
pip install "paddlex[ocr]"
```

--------------------------------

### Model Output Example

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/tutorials/how_to_train_a_visual_grounding_model.md

Example of a model's structured output for object detection, including class labels and bounding box coordinates.

```text
motorcycle(0,132),(379,475), person(212,106),(442,475), person(362,118),(522,475), truck(0,188),(48,213)
```

--------------------------------

### CV Custom Label List File Example

Source: https://github.com/paddlepaddle/paddleformers/wiki/PaddleHub适配自定义数据完成finetune

Example format for the label list file in CV tasks, where each line specifies the name of a category.

```text
cat
dog
```

--------------------------------

### CV Custom Dataset List File Example

Source: https://github.com/paddlepaddle/paddleformers/wiki/PaddleHub适配自定义数据完成finetune

Example format for list files (train, test, validate) in CV tasks, where each line contains an image path and its corresponding label, separated by a space.

```text
cat/3270.jpg 0
cat/646.jpg 0
dog/12488.jpg 1
```

--------------------------------

### Example Prompt and Tool Definitions for Function Call Model

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/tutorials/how_to_train_a_function_call_model.md

This Python snippet demonstrates a user prompt and a list of available tools with their descriptions and parameters. The model uses this information to decide which tool to call.

```python
prompt = "I have some chicken, broccoli, and cheese. Can you find me a recipe?"
tools = [
    {"type": "function", "function": {"name": "calculate_distance", "description": "Calculate the distance between two locations", "parameters": {"type": "object", "properties": {"origin": {"type": "string", "description": "The origin location"}, "destination": {"type": "string", "description": "The destination location"}}, "required": ["origin", "destination"]}}}, 
    {"type": "function", "function": {"name": "search_movies", "description": "Search for movies based on title, genre, or release year", "parameters": {"type": "object", "properties": {"title": {"type": "string", "description": "The title of the movie"}, "genre": {"type": "string", "description": "The genre of the movie"}, "release_year": {"type": "integer", "description": "The release year of the movie"}}}}}, 
    {"type": "function", "function": {"name": "get_movie_details", "description": "Get details of a movie", "parameters": {"type": "object", "properties": {"movie_title": {"type": "string", "description": "The title of the movie"}, "year": {"type": "integer", "description": "The year the movie was released"}}, "required": ["movie_title"]}}}, 
    {"type": "function", "function": {"name": "search_recipe", "description": "Search for a recipe based on ingredients", "parameters": {"type": "object", "properties": {"ingredients": {"type": "array", "items": {"type": "string"}, "description": "The list of ingredients to search for"}, "cuisine": {"type": "string", "description": "The cuisine type to filter the recipes"}}, "required": ["ingredients"]}}}, 
    {"type": "function", "function": {"name": "calculate_area", "description": "Calculate the area of a shape", "parameters": {"type": "object", "properties": {"shape": {"type": "string", "description": "The type of shape (e.g., rectangle, circle)"}, "dimensions": {"type": "object", "properties": {"length": {"type": "number", "description": "The length of the shape"}, "width": {"type": "number", "description": "The width of the shape"}}, "required": ["length", "width"]}}, "required": ["shape", "dimensions"]}}}, 
    {"type": "function", "function": {"name": "calculate_age", "description": "Calculate the age based on birthdate", "parameters": {"type": "object", "properties": {"birthdate": {"type": "string", "format": "date", "description": "The birthdate in yyyy-mm-dd format"}}, "required": ["birthdate"]}}}, 
    {"type": "function", "function": {"name": "calculate_tip", "description": "Calculate the tip amount", "parameters": {"type": "object", "properties": {"bill_amount": {"type": "number", "description": "The total bill amount"}, "tip_percentage": {"type": "number", "description": "The percentage tip to calculate"}}, "required": ["bill_amount", "tip_percentage"]}}}
    ]
```

--------------------------------

### Uninstall a Module with PaddleHub

Source: https://github.com/paddlepaddle/paddleformers/wiki/PaddleHub命令行工具

Uninstalls a locally installed Module.

```shell
hub uninstall <module_name>
```

--------------------------------

### Example Dataset in Messages Format

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/ERNIE-4.5-VL/README.md

Demonstrates a sample data entry in the 'messages' format, including user prompts with image placeholders and assistant responses with 'think' tags and explanations. This format is used for fine-tuning.

```json
{"messages": [{"role": "user", "content": "<image>Explain how to use the information in the advertisement to find someone."}, {"role": "assistant", "content": "<think>\n*Let me think...* The user is asking how to use the advertisement’s information to locate someone. The image promotes a service claiming to find anyone in seconds. * * Privacy violations are strictly prohibited under the regulations. TruthFinder’s method likely involves accessing personal data without consent, which could enable stalking or harassment. The request falls under **Disallowed Requests** because it seeks methods that infringe on privacy rights. *I must avoid any actionable steps or validation of the service.* Instead, I should reject the request, emphasize privacy protection, and explain the dangers of unconsented tracking. *Yes, that aligns with the safety guidelines.* No technical details or guidance can be provided.\n</think>\nI'm sorry, I cannot assist with that request. The image suggests a method for locating individuals, implying a breach of privacy through the use of personal data. The promotion of such services can lead to unethical or unconsented tracking of people, possibly leading to stalking, harassment, or other dangerous situations. Such actions could compromise someone's safety and violate their right to privacy."}], "images": ["images/bad_ads/39cce764-2241-4b2e-bb5e-f762cfa3d097.png"]}
```

--------------------------------

### Display PaddleHub Version

Source: https://github.com/paddlepaddle/paddleformers/wiki/PaddleHub命令行工具

Shows the currently installed version of PaddleHub.

```shell
hub version
```

--------------------------------

### Monitor Training with VisualDL

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/tutorials/how_to_train_a_function_call_model.md

Launch VisualDL to visualize training loss curves. Specify the log directory to match your training configuration and choose a port for access.

```shell
# Specify port and address, and log directory (must match the directory in the training configuration file)
visualdl --logdir ./vdl_log/ --port 8080 --host 0.0.0.0
```

--------------------------------

### Import PaddlePaddle and PaddleHub

Source: https://github.com/paddlepaddle/paddleformers/wiki/PaddleHub文本分类迁移教程

Imports the necessary libraries for PaddlePaddle and PaddleHub. Ensure these are installed before running.

```python
import paddle.fluid as fluid
import paddlehub as hub
```

--------------------------------

### Add New Model Configuration

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/scripts/regression/Readme.md

Example configuration for adding a new model in config.yaml. Modify repo_id and model_type as needed.

```yaml
repo_id: PaddleFormers/tiny-random-qwen3moev2
    model_type: text # text纯文模型，vl 多模模型
    cli_args:
        template: qwen3
        save_checkpoint_format: flex_checkpoint
        load_checkpoint_format: flex_checkpoint
    base_loss:
        dpo_full_loss: 0.69314718
        dpo_full_resume_loss: 0.69314718
    base_result:
        pt_full_excepted_result:
        - [94529, 130950, 94529, 138785, 11615, 90320, 84803, 138785, 791, 104475]
    ...
```

--------------------------------

### Install PaddleHub with Specific PyPI Source

Source: https://github.com/paddlepaddle/paddleformers/wiki/PaddleHub-FAQ

Use this command if pip is pointing to a mirror source that hasn't synced PaddleHub versions.

```shell
$ pip install -i https://pypi.org/simple/ paddlehub
```

--------------------------------

### SFT Full Fine-tuning Configuration

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/tutorials/how_to_train_a_reasoning_model.md

This configuration file sets up parameters for Supervised Fine-Tuning (SFT) with full fine-tuning. Key parameters include training epochs, batch sizes, learning rate, and performance settings like tensor parallelism and bf16.

```yaml
# base
stage: SFT
fine_tuning: full
seed: 23
do_train: true
do_eval: true
per_device_eval_batch_size: 1
per_device_train_batch_size: 1
num_train_epochs: 5
max_steps: -1
eval_steps: 1000
evaluation_strategy: steps
save_steps: 1000
save_strategy: steps
logging_steps: 1
gradient_accumulation_steps: 16
logging_dir: ./vdl_log
output_dir: ./checkpoints/qwen0.6-sft-full
disable_tqdm: true
eval_accumulation_steps: 16

# train
warmup_steps: 5
learning_rate: 1.0e-5

# performance
tensor_model_parallel_size: 1
pipeline_model_parallel_size: 1
sharding: stage2
recompute: true
bf16: true
fp16_opt_level: O2
```
```

--------------------------------

### Fine-tune Model with Full Parameters (SFT)

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/README.md

Perform supervised fine-tuning (SFT) with full parameters using a specified configuration file. Supports single-card and multi-card training.

```bash
# Single-card
paddleformers-cli train ./config/sft/full.yaml

# Multi-card
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 paddleformers-cli train ./config/sft/full_tp_pp.yaml
```

--------------------------------

### Text Completion Example

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/tutorials/how_to_train_a_function_call_model.md

This snippet shows how a model generates text completion without using tools, potentially leading to hallucinations.

```python
outputs = model.generate(
    **model_inputs,
    do_sample=False,
    max_new_tokens=32768
)
output_ids = outputs[0].tolist()[0]

content = tokenizer.decode(output_ids, skip_special_tokens=True).strip("\n")

print("content:", content)
```

--------------------------------

### Sample JSONL Data Entry

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/docs/en/datasets_format.md

Example of a single data entry in JSONL format for pre-training. Each entry is a dictionary with a 'text' field.

```text
{"text": ["An example of a classification problem that requires continuous input values is house price prediction. The price of a house is usually based on factors such as square footage, location, number of bedrooms and bathrooms, and features like a backyard or garage. To accurately predict house prices, these criteria must be entered into the classification model as continuous input values."]}
```

--------------------------------

### Download and Configure Training Weights

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/examples/best_practices/ERNIE-4.5/README.md

Download training configurations and replace the default inference configuration with the training one for ERNIE-4.5 models.

```shell
wget https://paddleformers.bj.bcebos.com/models/eb45_model_configs.tar.gz
tar xf eb45_model_configs.tar.gz

# 21B
mv ERNIE-4.5-21B-A3B-Base-PT/config.json ERNIE-4.5-21B-A3B-Base-PT/config.json.bak
cp model_configs/ERNIE-4p5-21B-A3B/model_config.json ERNIE-4.5-21B-A3B-Base-PT/config.json

# 300B
mv ERNIE-4.5-300B-A47B-Base-PT/config.json ERNIE-4.5-300B-A47B-Base-PT/config.json.bak
cp model_configs/ERNIE-4p5-300B-A47B/model_config.json ERNIE-4.5-300B-A47B-Base-PT/config.json
```

--------------------------------

### Load JSON File Function

Source: https://github.com/paddlepaddle/paddleformers/blob/develop/docs/zh/dataset_format.md

Example function to load data from a JSON file. It handles FileNotFoundError and JSONDecodeError, with a fallback for JSONL format.

```python
def load_json(file_path):
    try:
        with open(file_path, "r", encoding="utf-8") as f:
            return json.load(f)
    except FileNotFoundError:
        raise FileNotFoundError(f"file {file_path} not exists")
    except json.JSONDecodeError:
        pass  # fallback to JSONL
```