Configuration File (settings.jsonc)
WeClone reads settings.jsonc in the project root by default. Start from a copy of settings.template.jsonc:
cp settings.template.jsonc settings.jsoncTo use another file, run weclone-cli --config-path <path> <command> or set the WECLONE_CONFIG_PATH environment variable. The file may contain comments. If version differs from the config version the project expects, WeClone prints a warning at startup. Update your config against the template in that case, and see Upgrading from 0.3 for older versions.
Only the common settings are listed below. Defaults in the tables are the template values.
security_args
How data is stored. See Security.
| Setting | Default | Description |
|---|---|---|
storage_mode | "encrypted" | encrypted stores chat data and derived files encrypted, plaintext stores them as plain files. Fixed after first initialization. Changing it means processing all data again. Treated as plaintext when missing |
state_dir | ".weclone" | Directory for keys, tokens and assistant settings |
runtime_dir | System temp directory | Where decrypted temporary files go in encrypted mode |
common_args
Settings shared across features, mainly for fine-tuning and inference.
| Setting | Default | Description |
|---|---|---|
model_name_or_path | "./models/Qwen2.5-7B-Instruct" | Base model path. See Model Download |
adapter_name_or_path | "./model_output" | LoRA weights directory. Training writes to it and inference loads from it |
template | "qwen" | LLaMA Factory chat template. Must match the model |
default_system | "请你扮演一名人类，不要说自己是人工智能" | System prompt for training and inference. The system prompt on chat bot platforms must match it exactly |
media_dir | "dataset/media" | Directory for images and other media |
image_max_pixels | 409920 | Maximum image pixels in multimodal training. 409920 is about 720P |
finetuning_type | "lora" | Fine-tuning method. Only LoRA is supported |
enable_thinking | false | Whether to enable the model's thinking mode |
trust_remote_code | true | Whether to trust code in the model repository when loading |
cli_args
| Setting | Default | Description |
|---|---|---|
log_level | "INFO" | Terminal log level: DEBUG, INFO, WARNING, ERROR, CRITICAL. logs/weclone.log always records DEBUG and above |
full_log | false | Also write the terminal output of third-party libraries to logs/weclone.log. Plaintext mode only |
codex_exec_args
Settings for the Codex CLI backend, used by Build Your Portrait and Speaking Style.
| Setting | Default | Description |
|---|---|---|
model | "gpt-6.1-sol" | Model to use. Pick one at least as capable as gpt-6-sol medium |
effort | "medium" | Reasoning effort |
batch_size | 30 | Concurrent requests |
request_interval_seconds | 0.5 | Interval between request launches (seconds) |
capacity_cooldown_seconds | 20.0 | Pause before new requests when the model is busy or rate-limited (seconds) |
timeout | 300 | Timeout per request (seconds) |
command | "codex" | Codex executable. With several versions installed, give the absolute path of one that supports the model |
sandbox | "read-only" | Codex sandbox mode |
agent_distill_args
Settings for building the life portrait. See Build Your Portrait.
| Setting | Default | Description |
|---|---|---|
llm_provider | "codex_exec" | Model backend. codex_exec reads codex_exec_args. api uses base_url, llm_api_key and model_name in make_dataset_args. See Model backend |
batch_size | 10 | Concurrent requests for the API backend |
timeout | 120 | Timeout per request for the API backend (seconds) |
max_tokens | No limit | Output limit per request for the API backend |
max_samples_per_window | 8 | Short conversation segments are combined into one request. This is the most segments per request |
max_content_chars | 400 | Most message characters per request, not counting IDs, times and role labels. Longer segments get a request of their own |
Extraction and organizing results are cached by input in dataset/res_csv/agent/llm_cache.sqlite, so reruns only request what is missing. Delete the file to request everything again.
harness_args
Settings for the model-driven clone. This feature is in development. No changes are needed for now.
make_dataset_args
Settings for data processing and the fine-tuning dataset. See Data Preprocessing for make-dataset.
| Setting | Default | Description |
|---|---|---|
platform | "chat" | Data source: chat (generic CSV) or telegram. See Preparing Data |
language | "zh" | Main chat language, zh or en. Selects the recognizer used for privacy filtering |
telegram_args.my_id | "user1234567890" | Required for Telegram data: your own ID, the from_id of messages you sent in result.json |
include_type | ["text"] | Message types in the training data. Add "image" to train a multimodal model |
max_image_num | 2 | Most images per training example |
blocked_words | Example values | Blocked words. Messages containing one are left out of the training data. Words in dataset/blocked_words.json are used too |
add_time | false | Add message times to the training data |
add_relation | false | Add your relationship with the other person to the training data, taken from users.json |
single_combine_time_window | 2 | Consecutive messages from one person within this many minutes are combined into one |
qa_match_time_window | 5 | The other person's message and your reply within this many minutes form a QA pair |
combine_msg_max_length | 2048 | Maximum length of a combined message. Used with cutoff_len |
messages_max_length | 2048 | Maximum characters per training example. Used with cutoff_len |
prompt_with_history | false | Whether to include conversation history in the prompt. Has no effect on multimodal data |
clean_dataset | Off | Score training data with an LLM and filter it. See Data Preprocessing |
online_llm_clear | false | Clean with an online API instead of local vLLM |
base_url, llm_api_key, model_name | Example values | Address, key and model name of an OpenAI-compatible API. Used by online cleaning and by the API backend of the life portrait |
clean_batch_size | 50 | Batch size for online cleaning |
vision_api | Off | Turn images into text descriptions with a multimodal API. See Data Preprocessing |
train_sft_args
Fine-tuning parameters. See Model Fine-tuning. Any LLaMA Factory parameter can go here, see LLaMA Factory arguments. WeClone adds these:
| Setting | Description |
|---|---|
quantization | QLoRA settings, also used for inference. See Enable QLoRA |
resume_adapter_name_or_path | Continue training from existing LoRA weights. Output still goes to common_args.adapter_name_or_path |
deepspeed | DeepSpeed config file for multi-GPU training, such as "ds_config.json" |
infer_args
Sampling parameters for inference, used by the web demo, the API server and testing.
| Setting | Default | Description |
|---|---|---|
repetition_penalty | 1.2 | Repetition penalty |
temperature | 0.5 | Temperature |
max_length | 256 | Maximum generation length |
top_p | 0.65 | Top-p sampling |
vllm_args
vLLM parameters for offline data cleaning. See Notes on using vLLM.
| Setting | Default | Description |
|---|---|---|
gpu_memory_utilization | 0.9 | Share of GPU memory vLLM may use |
dtype | Auto | Compute precision. Use "float16" on GPUs without bfloat16 support, such as Tesla T4 and V100 |
quantization, load_format | Not set | Quantized loading, such as "bitsandbytes" |
test_model_args
| Setting | Default | Description |
|---|---|---|
test_data_path | "dataset/eval/test_data-zh.json" | Test questions for test-model. See Model Inference and Testing |
embedding_service_args
Used by features in development. No changes are needed for now.