Preparing adapters

Basevex accepts LoRA adapters trained on a supported base model. The adapter is checked before any deployment can use it. Validation reads the configuration and the safetensors header only — your tensors are never loaded or executed during validation.

Files to upload

Upload the two files individually. Archives (zip, tar) and pickle-based formats such as adapter_model.bin are not accepted.

A valid configuration

{
  "peft_type": "LORA",
  "base_model_name_or_path": "Qwen/Qwen3-8B",
  "r": 16,
  "lora_alpha": 32,
  "lora_dropout": 0.05,
  "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj"],
  "bias": "none",
  "task_type": "CAUSAL_LM"
}

Validation rules

An adapter is rejected if any of these is true:

• peft_type is not LORA.
• r (rank) is higher than the base model’s maximum rank, shown on the model in your dashboard.
• target_modules includes a module the base model doesn’t allow. Typical allowed modules are the attention and MLP projections (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj); the exact list is shown per model.
• modules_to_save is set (for example, a retrained embedding or output layer).
• use_dora is true.
• The adapter adds tokens to the vocabulary.
• A tensor uses a dtype other than F16, BF16 or F32.
• A tensor name doesn’t map to an allowed module and layer, or a lora_A/lora_B shape doesn’t match the rank and the base model’s dimensions.
• The safetensors header is larger than 16 MB, or its offsets don’t account for the file size exactly.

If base_model_name_or_path doesn’t match the base model you selected, you’ll see a warning rather than a rejection — local training paths are common. The shape checks still catch an adapter trained on a different architecture.

Statuses

Licensing

When you upload, you confirm you have the right to use and host the adapter and its training data, and that use complies with the base model’s license and our Acceptable Use Policy.