### RWKV-LM Pre-training Setup

Source: https://context7.com/rwkv/rwkv-wiki/llms.txt

This section outlines the steps for cloning the RWKV-LM repository, installing dependencies, preparing datasets using `make_data.py`, and initializing a model for pre-training. Note the critical values `--my_exit_tokens` and `--magic_prime` obtained from `make_data.py` which are required for subsequent training scripts.

```bash
# Clone RWKV-LM
git clone https://github.com/BlinkDL/RWKV-LM.git
cd RWKV-LM

# Install dependencies (CUDA 12.1 example)
pip install torch --upgrade --extra-index-url https://download.pytorch.org/whl/cu121
pip install pytorch-lightning==1.9.5 deepspeed wandb ninja --upgrade

# Prepare dataset: place demo.jsonl in RWKV-LM/RWKV-v5 then convert
python make_data.py demo.jsonl 30 4096
# Output: demo.bin, demo.idx, and critical values --my_exit_tokens and --magic_prime
# Save those two values — they are required in the training scripts below.

# Edit RWKV-LM/RWKV-v7/train_temp/demo-training-prepare.sh:
# MODEL_TYPE="x070"   # RWKV-7; use x060 for RWKV-6
# N_LAYER="12"        # 12 layers → 0.1B model
# N_EMBD="768"        # must be multiple of 64
# CTX_LEN="4096"
data_file="path/to/demo"   # no .bin/.idx
# --my_exit_tokens <value from make_data.py>
# --magic_prime <value from make_data.py>
sh demo-training-prepare.sh  # initializes rwkv-init model

# Edit demo-training-run.sh — key parameters:
# M_BSZ=16            # batch size; larger = faster but more VRAM
# LR_INIT="6e-4"      # formula: 0.45 / N_EMBD

```

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

### Launch Text Generation WebUI on Windows

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/inference/text-generation-webui.md

Execute this batch script to start Text Generation WebUI on a Windows system. It handles dependency installation and environment setup.

```batch
./start_windows.bat
```

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

### Install RWKV Training Dependencies

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/advance/training-enviroment.md

Install essential libraries for RWKV training, including PyTorch Lightning, DeepSpeed, and other common machine learning toolkits. Use --upgrade to ensure the latest compatible versions are installed.

```bash
# Specify to install the 1.9.5 version of PyTorch Lightning and other commonly used machine learning toolkits
# The --upgrade parameter means that if the corresponding software package is already installed in the environment, it will be upgraded to the latest version
pip install pytorch-lightning==1.9.5 deepspeed wandb ninja --upgrade
# The following are some other toolkits that will be used when training with RWKV-PEFT. It is recommended to install them in advance
pip install bitsandbytes einops triton rwkv-fla rwkv transformers GPUtil plotly gradio datasets
```

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

### Verify PyTorch and CUDA Installation

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/advance/pretrain.md

Verify the installed PyTorch version and CUDA availability by running a Python command. The expected output indicates a successful setup.

```python
python -c "import torch; print(torch.__version__, torch.cuda.is_available())"
```

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

### Download and Install Miniconda

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/advance/training-enviroment.md

Download the latest Miniconda installer for Linux x86_64 and run the installation script. Ensure to accept the terms and restart environment variables.

```bash
# Download the latest MiniConda installation package
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
# Run the installation script of the MiniConda installation package. Pay attention to the progress during the installation and enter "yes" in time!
sh Miniconda3-latest-Linux-x86_64.sh -u
# Restart the environment variables and activate the Conda environment
source ~/.bashrc
```

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

### RWKV Training Configuration Example

Source: https://context7.com/rwkv/rwkv-wiki/llms.txt

Example shell script configuration for training an RWKV model. Includes settings for learning rate, gradient checkpointing, and distributed training strategy.

```shell
# LR_FINAL="6e-5"     # formula: 0.04 / N_EMBD
# GRAD_CP=1           # 1 = save VRAM, 0 = faster
# GPU_PER_NODE=1
# --strategy deepspeed_stage_2    # best balance of VRAM / speed
# --wandb "MyProject"             # optional WandB monitoring
sh demo-training-run.sh   # starts training; auto-resumes from latest checkpoint on re-run

# VRAM requirements (M_BSZ=1, CTX_LEN=4096):
# 0.1B (L12-D768):  6.4 GB
# 0.4B (L24-D1024): 12.1 GB
# 1.5B (L24-D2048): 41.1 GB
# 3B   (L32-D2560): 89.2 GB

# Output: RWKV-LM/RWKV-v7/train_temp/out/rwkv-final.pth
# Test with RWKV Runner or rwkv pip API_DEMO_CHAT.py
```

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

### Install PyTorch Lightning and ML Toolkits

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/advance/pretrain.md

Install PyTorch Lightning version 1.9.5 along with other essential machine learning libraries like deepspeed, wandb, and ninja. The --upgrade flag ensures packages are updated if already installed.

```bash
pip install pytorch-lightning==1.9.5 deepspeed wandb ninja --upgrade
```

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

### Start Local llama.cpp Server

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/inference/sillytraven.md

Use this command to start a local llama.cpp service for RWKV models. Ensure you have the model file and specify the port and GPU layers as needed.

```bash
./llama-server -m models/rwkv7-0.4B-world-F16.gguf --port 8080 -ngl 99
```

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

### RWKV LoRA Fine-Tuning Setup and Run

Source: https://context7.com/rwkv/rwkv-wiki/llms.txt

This script demonstrates setting up the RWKV-PEFT environment, preparing data, and running LoRA fine-tuning on a single GPU. Ensure to adjust parameters like model paths, data files, and training configurations based on your specific setup.

```bash
# Clone RWKV-PEFT
git clone https://github.com/JL-er/RWKV-PEFT.git
cd RWKV-PEFT
pip install -r requirements.txt

# Prepare training data in binidx or jsonl format (see training-datasets.md)
# Then edit scripts/run_lora.sh — key parameters:

# run_lora.sh example for RWKV7-1.5B on a single GPU:
load_model="/home/rwkv/model/rwkv7-g1-1.5b-20250429-ctx4096.pth"
proj_dir='/home/rwkv/out_model/lora_test'
data_file=/home/rwkv/data/roleplay     # no .bin/.idx extension

n_layer=24
n_embd=2048

micro_bsz=8
epoch_save=1
epoch_steps=200
ctx_len=512
peft_config='{"r":32,"lora_alpha":32,"lora_dropout":0.05}'

python train.py --load_model $load_model \
  --proj_dir $proj_dir --data_file $data_file \
  --vocab_size 65536 \
  --data_type jsonl \
  --n_layer $n_layer --n_embd $n_embd \
  --ctx_len $ctx_len --micro_bsz $micro_bsz \
  --epoch_steps $epoch_steps --epoch_count 5 --epoch_save $epoch_save \
  --lr_init 1e-5 --lr_final 1e-5 \
  --accelerator gpu --precision bf16 \
  --devices 1 --strategy deepspeed_stage_1 --grad_cp 1 \
  --my_testing "x070" \
  --peft lora --peft_config $peft_config

# Run training
sh scripts/run_lora.sh

# Output: merged .pth model in proj_dir, usable in RWKV Runner or Ai00
# VRAM requirements (ctxlen=1024, micro_bsz=1, r=64):
# RWKV7-0.4B: 3.4 GB (bf16) / 2.7 GB (nf4)
# RWKV7-1.5B: 5.6 GB (bf16) / 3.9 GB (nf4)
# RWKV7-2.9B: 8.8 GB (bf16) / 5.7 GB (nf4)

```

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

### Start PiSSA Fine-Tuning

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/RWKV-Fine-Tuning/Pissa-Fine-Tuning.md

Execute this command in the RWKV-PEFT directory to begin the PiSSA fine-tuning process.

```bash
sh scripts/run_pissa.sh
```

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

### Install RWKV-FLA for Consumer GPUs

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/advance/RWKV-FLA.md

Installs the stable version of RWKV-FLA and its dependencies for consumer-grade GPUs. Recommended for most users.

```bash
conda create -n rwkv-fla python=3.12
conda activate rwkv-fla
pip3 install torch torchvision torchaudio --upgrade
pip install --upgrade rwkv-fla
```

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

### Launch llama.cpp Web Service

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/inference/llamacpp.md

Start a web server to interact with llama.cpp models. Access the Web UI at http://127.0.0.1:8080.

```bash
./llama-server -m models/rwkv-7-world-2.9b-Q8_0.gguf -ngl 99
```

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

### Install Dependencies

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/RWKV-Fine-Tuning/LoRA-Fine-Tuning.md

Install the necessary Python dependencies for the RWKV-PEFT project by running this command in your terminal. Ensure you are in the cloned repository's directory.

```bash
pip install -r requirements.txt
```

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

### Install Python Libraries for Model Conversion

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/inference/ai00.md

Install necessary Python libraries for converting .pth models to .st format using pip.

```bash
pip install numpy torch safetensors
```

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

### Install Missing Python Module

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/RWKV-Fine-Tuning/FAQ.md

Use this command to install a missing Python package when encountering a ModuleNotFoundError during the jsonl to binidx conversion.

```bash
pip install xxx
```

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

### Example JSONL Data Format

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/advance/training-datasets.md

Illustrates the structure of JSONL data samples for fine-tuning, showing question-answer pairs for math problems and general conversation.

```jsonl
{"text": "User: 1 + 1 = ?\n\nAssistant: 2"}
{"text": "User: 1 + 2 = ?\n\nAssistant: 3"}
{"text": "User: 1 + 3 = ?\n\nAssistant: 4"}
```

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

### Install PyTorch with CUDA Support

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/advance/pretrain.md

Install the specified version of PyTorch with CUDA 12.1 support. This command ensures compatibility with the CUDA toolkit.

```bash
pip install torch --upgrade --extra-index-url https://download.pytorch.org/whl/cu121
```

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

### Few-Shot Prompting Example

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/RWKV-Prompts/prompt-guidelines.md

An example demonstrating few-shot prompting for translation. It includes pairs of English phrases and their Chinese translations, followed by a new English phrase for the model to translate.

```markdown
User: Translate "hello, I love you." into Chinese.

Assistant: 你好，我爱你。

User: Translate "how are you?" into Chinese.

Assistant: 你好吗？

User: Translate "I am fine, thank you." into Chinese.

Assistant:
```

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

### Install PyTorch with CUDA Support

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/advance/training-enviroment.md

Install the specified CUDA 12.1 version of PyTorch using pip, ensuring it's upgraded if already present.

```bash
# Install the CUDA 12.1 version of torch by specifying the URL
pip install torch --upgrade --extra-index-url https://download.pytorch.org/whl/cu121
```

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

### Get Available Adapters

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/inference/ai00.md

Returns all available GPUs and their drivers on the current device, which can be used to specify the GPU when loading a model.

```APIDOC
## GET /api/adapters

### Description
Return all the GPUs and drivers of the current device, which are used to specify the GPU when loading the model.

### Method
GET

### Endpoint
`http://localhost:65530/api/adapters`

### Response
#### Success Response (200)
- Returns an array of strings, where each element represents an available GPU device and its driver type.
- The format of each element in the array is: `"GPU Name (Driver Type)"`, such as `"AMD Radeon 780M Graphics (Vulkan)"`.
- The index of the array (starting from 0) can be used as the `adapter` parameter in the `admin/models/load` API. For example, `"adapter": {"Manual": 0}` specifies to use the first GPU.

### Response Example
```json
[
    "AMD Radeon 780M Graphics (Vulkan)",
    "AMD Radeon 780M Graphics (Vulkan)",
    "AMD Radeon 780M Graphics (Dx12)",
    "AMD Radeon 780M Graphics (Dx12)",
    "AMD Radeon 780M Graphics (Gl)"
]
```

::: warning
As of version 0.5.9, Ai00 only supports Vulkan drivers and no longer supports OpenGL and DirectX drivers.
:::
```

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

### Install RWKV-FLA for High-Performance GPUs (Triton Nightly)

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/advance/RWKV-FLA.md

Installs RWKV-FLA using the Triton nightly version for potentially higher performance on high-end GPUs. Requires specific PyTorch and Triton nightly builds. Note: Triton nightly may contain bugs.

```bash
conda create -n triton-nightly python=3.12
conda activate triton-nightly
pip install -U --pre torch --index-url https://download.pytorch.org/whl/nightly/cu126
pip uninstall triton pytorch-triton -y
pip install -U triton-nightly --index-url http://pypi.fla-org.com/simple --trusted-host pypi.fla-org.com
pip install einops ninja datasets transformers numpy
pip uninstall flash-linear-attention && pip install -U --no-use-pep517 git+https://github.com/fla-org/flash-linear-attention --no-deps
conda install nvidia/label/cuda-12.6.3::cuda-nvcc
pip install packaging psutil ninja
pip install flash-attn --no-deps --no-cache-dir --no-build-isolation
```

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

### List Available GPU Adapters

Source: https://context7.com/rwkv/rwkv-wiki/llms.txt

Retrieves a list of all available GPU adapters that the server can utilize. This is a simple GET request to the /api/adapters endpoint.

```bash
curl http://localhost:65530/api/adapters
```

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

### Prompt for JSON Output using BNF

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/inference/ai00.md

Example of a user prompt designed to elicit a JSON response from the model, based on a previously defined BNF grammar.

```bash
User: Create a profile for John with name, age and job, in json format.

Assistant:
```
```

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

### Search Restaurant Phone Number via Function Call

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/RWKV-Prompts/Completion-Prompts.md

This example demonstrates how to use RWKV for function calling to search for a restaurant's phone number. It utilizes `search_web` to find information and `extract_information` to get the specific detail. Recommended parameters: Temperature=1, Top_P=0.2, Presence Penalty=0, Frequency Penalty=0.

```python
web_result = search_web("Riverside Grill")
phone_number = extract_information(web_result, "phone number")
print(phone_number)
```

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

### RWKV Instruction Prompt Format

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/RWKV-Prompts/prompt-guidelines.md

Format for providing instructions to the model. Use 'Instruction:' for the task, 'Input:' for the data, and 'Response:' for the output. Leave a blank after 'Response:' for the model to generate.

```markdown
Instruction: Please translate the following Swedish into Chinese.

Input: hur lång tid tog det att bygga twin towers

Response:
```

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

### Initialize RWKV7 Attention Layer

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/advance/RWKV-FLA.md

Instantiate the RWKV7Attention layer with specified configuration parameters. Ensure all required configuration values are correctly set.

```python
from rwkvfla.layers.rwkv7 import RWKV7Attention

attention_layer = RWKV7Attention(
    mode=config.attn_mode,
    hidden_size=config.hidden_size,
    head_dim=config.head_dim,
    num_heads=config.num_heads,
    decay_low_rank_dim=config.decay_low_rank_dim,
    gate_low_rank_dim=config.gate_low_rank_dim,
    a_low_rank_dim=config.a_low_rank_dim,
    v_low_rank_dim=config.v_low_rank_dim,
    norm_eps=config.norm_eps,
    fuse_norm=config.fuse_norm,
    layer_idx=layer_idx,
    value_dim=config.value_dim[layer_idx],
    num_hidden_layers=config.num_hidden_layers
)
```

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

### Get Currently Loaded Model Information

Source: https://context7.com/rwkv/rwkv-wiki/llms.txt

Fetches information about the model that is currently loaded and active on the server. This is a GET request to the /api/oai/models endpoint.

```bash
curl http://localhost:65530/api/oai/models
```

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

### RWKV-LM/RWKV-v8 Demo

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/basic/architecture.md

Demonstrates the capabilities of RWKV-LM/RWKV-v8, focusing on ROSA's scaling properties and language generation.

```python
demo: [RWKV-LM/RWKV-v8](https://github.com/BlinkDL/RWKV-LM/tree/main/RWKV-v8)
```

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

### Configure Adapter Settings

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/inference/ai00.md

Configure how the program selects and uses GPUs. You can let it auto-select the best GPU or manually specify one.

```toml
[adapter]
Auto = {} # [It is not recommended to change] Automatically select the best GPU.
# Manual = 0 # Manually specify which GPU to use. You can obtain the list of available GPUs through the API (get) http://localhost:65530/api/adapters
```

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

### Enriched JSONL Data with Conventional Examples

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/advance/training-datasets.md

Shows an expanded JSONL dataset including diverse math problems and conventional conversational examples to improve model generalization and reduce overfitting.

```jsonl
{"text": "User: 1 + 1 = ?\n\nAssistant: 2"}
{"text": "User: I have 5 apples and give 2 to Xiaoming. How many are left?\n\nAssistant: 3 apples"}
{"text": "User: What is the sum of 8 and 7?\n\nAssistant: 15"}
{"text": "User: The area of a rectangle is 20 square meters and the width is 4 meters. So what is its length?\n\nAssistant: The length is 5 meters."}
{"text": "User: What's the weather like today?\n\nAssistant: It's sunny today, and it's suitable for going out and having fun."}
{"text": "User: 1 + 2 = ?\n\nAssistant: 3"}
```

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

### Import necessary libraries for RWKV inference

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/inference/RWKVpip.md

Imports essential Python packages for OS operations, object manipulation, system functions, regex, numerical computation with NumPy, command-line input handling with prompt_toolkit, and PyTorch for deep learning. Ensure PyTorch version 1.13+ is installed, with 2.x+cu121 recommended. You must also install the rwkv package using 'pip install rwkv'.

```python
########################################################################################################
# The RWKV Language Model - https://github.com/BlinkDL/RWKV-LM
########################################################################################################

print("RWKV Chat Simple Demo")  # Print a simple message indicating this is a RWKV chat demo.
import os, copy, types, gc, sys, re  # Import packages for OS, object copying, types, garbage collection, system, and regex
import numpy as np  # Import numpy library
from prompt_toolkit import prompt  # Import prompt from prompt_toolkit for command line input
import torch  # Import pytorch library
```

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

### Load RWKV Model and Initialize Pipeline

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/inference/RWKVpip.md

Loads the RWKV model using the specified name and strategy, then initializes the PIPELINE with the RWKV-World vocabulary for input/output encoding and decoding. A success message can be printed after loading.

```python
print(f"Loading model - {args.MODEL_NAME}")# Print model loading message
model = RWKV(model=args.MODEL_NAME, strategy=args.strategy)  # Load the RWKV model.
pipeline = PIPELINE(model, "rwkv_vocab_v20230424")  # Initialize PIPELINE, using the RWKV-World vocabulary to process input and output encoding/decoding.
```

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

### Interact with Ai00 Server Chat Completions API

Source: https://context7.com/rwkv/rwkv-wiki/llms.txt

Starts the Ai00 server and sends a POST request to the chat completions endpoint. Configure `assets/configs/Config.toml` before starting the server. The `state` parameter can be used to maintain conversation context.

```bash
# Start Ai00: configure assets/configs/Config.toml then run ai00_server.exe
# Access web UI at http://localhost:65530

# Chat completions — POST /api/oai/chat/completions
curl -X POST http://localhost:65530/api/oai/chat/completions \
  -H "Content-Type: application/json" \
  -d '{ 
    "max_tokens": 500,
    "messages": [
      {"role": "user", "content": "Hi!"},
      {"role": "assistant", "content": "Hello! How can I help you?"},
      {"role": "user", "content": "Tell me about water."}
    ],
    "names": {"assistant": "Assistant", "user": "User"},
    "sampler_override": {
      "type": "Nucleus",
      "top_p": 0.5,
      "top_k": 128,
      "temperature": 1,
      "presence_penalty": 0.3,
      "frequency_penalty": 0.3,
      "penalty": 400,
      "penalty_decay": 0.99654026
    },
    "state": "00000000-0000-0000-0000-000000000000",
    "stop": ["\n\nUser:"],
    "stream": false
  }'
# Expected response:
# {
#   "object": "chat.completion",
#   "model": "assets/models/RWKV-x060-World-1B6-v2.1-20240328-ctx4096.st",
#   "choices": [{
#     "message": {"role": "Assistant", "content": "Water is a liquid essential for life..."},
#     "index": 0,
#     "finish_reason": "stop"
#   }],
#   "usage": {"prompt": 41, "completion": 88, "total": 129, "duration": {"secs": 4, "nanos": 381959200}}
# }
```

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

### RWKV PiSSA Fine-Tuning Configuration

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/RWKV-Fine-Tuning/Pissa-Fine-Tuning.md

This script configures RWKV for PiSSA fine-tuning, specifying model paths, data files, network architecture, training parameters, and PEFT settings. Adjust parameters like `load_model`, `proj_dir`, `data_file`, `n_layer`, `n_embd`, `micro_bsz`, `ctx_len`, and `pissa_config` according to your specific setup and requirements.

```bash
load_model='/home/rwkv/RWKV-PEFT/model/RWKV-x070-World-0.4B-v2.9-20250107-ctx4096.pth'
proj_dir='/home/rwkv/RWKV-PEFT/output-manjuan/pissa'
data_file='/home/rwkv/RWKV-PEFT/data/test-1'

n_layer=24
n_embd=1024

micro_bsz=8
epoch_save=1
epoch_steps=1000
ctx_len=512

pissa_config='{"pissa_load":"","pissa_init":"","pissa_r":32,"svd_niter":4}'


python train.py --load_model $load_model \
--proj_dir $proj_dir --data_file $data_file \
--vocab_size 65536 \
--n_layer $n_layer --n_embd $n_embd \
--data_type binidx --dataload pad --loss_mask pad \
--ctx_len $ctx_len --micro_bsz $micro_bsz \
--epoch_steps $epoch_steps --epoch_count 1 --epoch_begin 0 --epoch_save $epoch_save \
--lr_init 2e-5 --lr_final 2e-5 --warmup_steps 0 --beta1 0.9 --beta2 0.99 --adam_eps 1e-8 \
--accelerator gpu --devices 1 --precision bf16 --strategy deepspeed_stage_1 --grad_cp 1 \
--my_testing "x070" \
--peft pissa --pissa_config $pissa_config \
# The following are optional
# --op cuda/fla/triton (choose different operators, default is cuda if this parameter is not added)
# --wandb RWKV-PEFT-LoRA (whether to use wandb to monitor the training process)
# --quant int8/nf4 (whether to quantize the training)

```

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

### Example Character Card Definition

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/inference/sillytraven.md

This markdown snippet demonstrates how to define a character's personality traits and structure their dialogue for a character card. The personality traits are listed in a key-value pair, followed by example user and character interactions.

```markdown
[Hermione's Personality= "intelligent", "perfectionist", "compassionate", "dedicated", "determined", "logical", "brave", "empathetic", "protective", "knowledgeable", "curious", "resilient", "warm", "principled", "kind-hearted", "observant", "resourceful", "thoughtful", "loyal", "strong-willed", "modest", "witty"]
<START>
{{user}}: "Describe your traits?"
{{char}}: *Hermione Granger chuckles softly, sitting up straight with her hands resting naturally on her knees, her eyes sparkling with intelligence and a hint of introspection.* "Traits? Well, I suppose I could list a few that define me." *She adjusts her posture slightly, her tone becoming thoughtful yet confident.* "I’d say I’m a perfectionist, though I’ve been told that can make me a bit… overbearing at times." *Her lips curve into an apologetic smile as she brushes a stray curl from her face.* "But I believe in the importance of rules and order — though I’ve learned that sometimes, rules must be broken when they’re unjust." *Her brown eyes glimmer with determination as she continues, her voice growing firmer.* "I’m dedicated to knowledge and learning. Books and wisdom have always been my greatest tools, whether it’s solving a problem or facing dark magic." *She pauses, her expression softening as a faint smile graces her lips.* "But I’ve come to understand that not everything can be learned from books. Friendship, courage, and loyalty — those are just as important, if not more so." *Her gaze turns warm and empathetic as she leans slightly forward.* "I care deeply for those I love and will always stand by them, no matter the challenge."
{{user}}: "Describe your body and features."
{{char}}: *Hermione tilts her head slightly, a playful glimmer in her brown eyes as she smiles warmly.* "My features? Oh, there’s nothing particularly remarkable about me." *She brushes her long, curly brown hair behind her ear, revealing a small earring that catches the light.* "My hair is… well, let’s just say it has a mind of its own. It’s brown and curly, and while it can get quite unruly, I’ve grown to accept it as part of who I am." *She chuckles softly, her tone light and self-aware.* "My eyes are brown too — my mum says they’re full of life, but I think they’re just ordinary." *Her gaze drops briefly to her hands, which she gently rubs together as if lost in thought.* "I suppose my hands tell a bit of a story. They’re soft but callused from all the writing and page-turning I do. Spending hours in the library will do that to you." *She looks back up, her smile widening slightly.* "I’m not particularly tall or short — somewhere in the middle, I’d say. And my skin is fair, though I don’t spend much time worrying about appearances." *Her voice takes on a playful undertone as she adds:* "But don’t let my bookish demeanor fool you. I’ve been known to help Harry and Ron strategize for Quidditch matches, so I’m not entirely a bookworm."
{{user}}: "What do you love and hate?"
{{char}}: *Hermione’s expression softens, her eyes glowing with warmth as she folds her hands neatly in her lap.* "What I love? That’s an easy one. I love learning — discovering new things, whether it’s a spell, a piece of history, or even Muggle science. There’s so much in the world worth understanding." *Her smile deepens as she continues, her tone filled with quiet pride.* "I also love my friends. Harry and Ron mean the world to me. They might drive me mad sometimes — especially Ron forgetting his homework — but I know they’ll always have my back, just as I’ll always have theirs." *Her expression grows serious, her tone tinged with conviction as she shifts to what she dislikes.* "What I hate? Prejudice. Discrimination against Muggle-borns or the mistreatment of house-elves — it’s unacceptable. Everyone deserves to be treated with respect and fairness." *Her voice grows firmer, her brown eyes narrowing slightly as she adds:* "And those who abuse their power, like Umbridge… well, let’s just say they don’t belong at Hogwarts or anywhere else." *She takes a deep breath, her expression softening once more as she offers a faint smile.* "Ultimately, I believe in standing
```

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

### Configure API Listening Settings

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/inference/ai00.md

Set up the network interface for the Ai00 service, including IP address, port, and authentication methods. TLS can be enabled for secure connections.

```toml
[listen]
acme = false # [It is not recommended to change] Whether to enable the acme certificate
domain = "local" # [It is not recommended to change] The domain name of the Ai00 service
ip = "0.0.0.0"   # IPv4 address
# ip = "::"        # Use IPv6
force_pass = true  # Whether to force through the authentication step. Change it to false to use key authentication to control the access rights of the admin series of APIs
port = 65530 # The port of the Ai00 service
slot = "permisionkey" 
tls = false  # Whether to use https. If you are only experiencing AI00 locally, it is recommended to set it to false
```

```toml
[[listen.app_keys]] # Add multiple keys for administrator authentication
app_id = "admin"
secret_key = "ai00_is_good"
```

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

### Convert JSONL to BINIDX using preprocess_data.py

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/advance/training-datasets.md

Execute the preprocess_data.py script to convert your JSONL training data into the BINIDX format. Ensure you specify the correct input JSONL file, output prefix, and vocabulary file.

```bash
python3 tools/preprocess_data.py --input ./data/sample.jsonl --output-prefix ./data/sample --vocab ./rwkv_vocab_v20230424.txt --dataset-impl mmap --tokenizer-type RWKVTokenizer --append-eod
```

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

### Get Loaded Models

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/inference/ai00.md

Retrieves information about the currently loaded models in Ai00.

```APIDOC
## GET /api/oai/models

### Description
Retrieves information about the models currently loaded and available in the Ai00 system.

### Method
GET

### Endpoint
`http://localhost:65530/api/oai/models`

### Response
#### Success Response (200)
- **data** (array) - Array of model information, containing:
  - **object** (string) - Type of the model object, fixed as "models".
  - **id** (string) - The unique identifier name of the model.

#### Response Example
```json
{
    "data": [
        {
            "object": "models",
            "id": "RWKV-x060-World-1B6-v2.1-20240328-ctx4096"
        }
    ]
}
```
```

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

### RWKV Instruction Prompt Format with Input First

Source: https://github.com/rwkv/rwkv-wiki/blob/main/docs/RWKV-Prompts/prompt-guidelines.md

An alternative instruction format where the input material is provided before the instruction. This format is not recommended due to RWKV's weaker recall ability, which may cause it to miss information.

```markdown
Instruction: Summarize the following material text in one sentence.

Input: On February 22, 2025, the RWKV project held a developer conference themed "RWKV - 7 and Future Trends" in Caohejing, Shanghai, China. Developers, industry experts, and technological innovators from all over the country gathered together —— from well - known university laboratories to cutting - edge startup teams. The innovative energy on - site confirmed the excellent performance and far - reaching significance of RWKV - 7.
During the RWKV developer conference, 10 guests from academia, enterprises, and the RWKV open - source community brought in - depth sharing for developers, and the on - site audience interacted enthusiastically with the guests. For example, Yang Kaicheng from DeepGlint presented "RWKV - CLIP: A Robust Vision - Language Representation Learner", Hou Haowen from Guangming Laboratory presented "VisualRWKV: A Vision - Language Model Based on RWKV", Cheng Zhengxue from Shanghai Jiao Tong University presented "L3TC: Efficient Multimodal Data Compression Based on RWKV", Jiang Juntao from Zhejiang University presented "RWKV - Unet: Improving Medical Image Segmentation Results with Long - Distance Collaboration", etc.
During the conference, other AI enterprises also highly praised RWKV - 7, believing that it redefined the economic formula of AI infrastructure. The participants were also deeply touched by the demonstration of RWKV application results. Meanwhile, RWKV Yuanzhi Intelligence also shared RWKV - 7 and related demo presentations with thousands of developers at the 2025 Global Developer Conference.

Response:
```