MDD-LLM-Llama3.2-1B-L2-ARCTIC

MDD-LLM with Llama 3.2 1B for L2-ARCTIC

This repository is a standalone IF-MDD inference bundle. It includes the complete model.ckpt inference state, the verified 44-label encoder, bundle-local Llama tokenizer and WavLM configuration, and all runtime modules needed by the wrapper, so loading does not contact Hugging Face.

Download

python -m pip install -U huggingface_hub
hf download Haopeng/MDD-LLM-Llama3.2-1B-L2-ARCTIC --local-dir MDD-LLM-Llama3.2-1B-L2-ARCTIC
cd MDD-LLM-Llama3.2-1B-L2-ARCTIC

Single WAV

pip install -r requirements.txt
python custom_interface.py \
  --hparams-file hyperparams.yaml \
  --audio /path/to/audio.wav \
  --override 'prompt_system_text=You are a pronunciation evaluator.' \
  --override 'prompt_user_text=Return only the perceived phoneme sequence.'

Repository Eval

With the main branch:

python train.py hparams/MDD_LLM_Llama3_2_1B.yaml \
  --mode eval \
  --inference_ckpt /path/to/MDD-LLM-Llama3.2-1B-L2-ARCTIC

Set HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 to enforce offline loading. manifest.json records the required recoverables, architecture signature, defaults, and label-encoder SHA256.