WeClone

English

简体中文

English

简体中文

Appearance

FAQ ​

TIP

When you run into a problem, try the following step by step:

  • Pull the latest code and update the environment: uv pip install -e . for the life portrait, uv pip install --group main -e . for fine-tuning. ​
  • If you upgraded from 0.3, check your configuration against Upgrading from 0.3. ​
  • Check the Issues and Discussions of the WeClone repository for similar problems. ​
  • Use DeepWiki to ask questions about the source code. ​

If the problem remains, open an issue.

Life portrait and assistants ​

specify --max-context-tokens explicitly ​

With the API backend (agent_distill_args.llm_provider set to api), WeClone can't know the model's context length. Add --max-context-tokens <input limit> to build-profile. The Codex backend asks for it too when it can't read the context length.

"结果目录非空" (the result directory is not empty) ​

--output-dir must be a new directory. Rerun into a new one, and content already extracted is read from dataset/res_csv/agent/llm_cache.sqlite instead of being requested again.

The clone purpose asks for ltp ​

Speaking style extraction needs the ltp package and the LTP model, see Installation. For a personal assistant only, choose --purpose assistant, which doesn't need them.

The Codex backend fails ​

Make sure codex runs in your terminal and is signed in, and that codex_exec_args.model is a model your Codex supports. With several Codex versions installed, set codex_exec_args.command to the absolute path of the right one.

Can I use English chats? ​

The portrait's extraction prompts are currently in Chinese, and English chats have not been tested yet.

Assistants can't read the portrait ​

  • Make sure weclone-cli server is still running.
  • In encrypted mode, turn on Allow assistants to read on the Assistants page. Turn it on again when the time is up or after the server restarts.
  • Assistants only get summaries approved on the Summaries page. After the portrait changes, the affected summaries have to be regenerated and approved.
  • Regenerating the token stops the old one at once, so every assistant has to be set up again with the new command.

I forgot my password ​

Run weclone-cli server-reset-password. The old encrypted data can't be recovered and has to be built again from the original chats in a new output directory, see Security.

Fine-tuned clone ​

Out of GPU memory ​

  • Adjust the training parameters
    1. Reduce the batch size: per_device_train_batch_size: 1
    2. Reduce the maximum sequence length: cutoff_len: 512
    3. Replace model operators: enable_liger_kernel: true and use_unsloth_gc: true
    4. Use DeepSpeed ZeRO-3 or FSDP to split model weights across devices, or CPU offloading (encrypted mode doesn't support distributed training yet)
    5. Fine-tune a quantized model (the weights can't be merged into the original model)
    6. Set quantization_bit: 4 under train_sft_args.quantization to use QLoRA (LoRA only), see Enable QLoRA
  • Rent a cloud GPU for fine-tuning, then download the model_output folder and run inference locally.
  • For multimodal models, reduce image_max_pixels and max_image_num.

The fine-tuned clone doesn't sound like me ​

  • Adjust fine-tuning parameters such as lora_rank and lora_dropout.
  • Use a larger model and more chat data.
    • 7B models give average results, 14B is acceptable and 32B is better.
  • Fine-tune a multimodal model so fewer conversations are cut.
  • Enable data cleaning.

"安全模式不能加载旧明文个人模型或 adapter" (encrypted mode can't load an old plaintext model) ​

Encrypted mode only loads models saved encrypted. After switching from plaintext mode, build the training data again and retrain.

LLaMA-Factory questions ​

Windows ​

  • Set "dataloader_num_workers": 0 in train_sft_args to avoid Can't pickle local object on Windows.

Chat bots ​

Bot requests return 401 ​

The API key must be the token generated on the Bots page of the web UI. Regenerating or revoking it stops the old token at once, and a password reset requires a new token too.

Bot requests return 403 ​

The /v1 API refuses requests that carry an Origin header, which browsers send. Call it from the bot platform's server.

A bot platform in Docker can't reach the server ​

The server listens only on 127.0.0.1 by default. On Linux, start it with --host 172.17.0.1 and use http://172.17.0.1:8005/v1 on the platform. On Docker Desktop (macOS, Windows), use http://host.docker.internal:8005/v1 on the platform. See Server and Bot Token.

Data ​

The chat CSV file shows garbled characters ​

Don't open it in Excel. Use an editor such as VS Code. In encrypted mode, export .enc files to plaintext with weclone-cli decrypt first.