PALATE-LoRA

This repository contains the five anonymous per-user LoRA adapters used as PALATE user simulators. All adapters share the post-trained Qwen/Qwen3.5-35B-A3B base model and the same rank and target-module family. The base-model weights are not included.

Related resources:

• Code: Zhuyh1139/PALATE
• Dataset and frozen rubrics: muset-ai/PALATE

U1/
  adapter_config.json
  adapter_model.safetensors
...
U5/
  adapter_config.json
  adapter_model.safetensors
checksums.json

Download

pip install -U huggingface_hub
hf download muset-ai/PALATE-LoRA --local-dir adapters/PALATE-LoRA

The PALATE GitHub tool then loads the local directories:

export PALATE_LORA_U1=adapters/PALATE-LoRA/U1
export PALATE_LORA_U2=adapters/PALATE-LoRA/U2
export PALATE_LORA_U3=adapters/PALATE-LoRA/U3
export PALATE_LORA_U4=adapters/PALATE-LoRA/U4
export PALATE_LORA_U5=adapters/PALATE-LoRA/U5
bash scripts/serve_loras.sh

For direct PEFT loading, download the snapshot and pass the selected local subdirectory to PeftModel.from_pretrained.

Training summary

• Base model: Qwen/Qwen3.5-35B-A3B
• Method: one LoRA per anonymous user
• Rank / alpha / dropout: 16 / 32 / 0.05
• Epochs: 1
• Learning rate: 1e-4
• Scheduler: cosine with 3% warmup
• Maximum sequence length: 8192

Training examples are deterministically derived from the annotated PALATE dataset by the public palate prepare-sft command.

Limitations and use policy

These adapters simulate interaction behavior from a small research cohort. They are not representations of demographic groups and must not be used for identity inference, impersonation, surveillance, or consequential profiling. See MODEL_USE_POLICY.md.

The adapters are released under CC BY-NC 4.0. The base model remains governed by its own license.