WillHeld/DiVA-llama-3-token-align-8b

Model Card for Diva Llama 3

This is an ablation of our Distilled Voice Assistant (DiVA) model which can handle speech and text as inputs. This ablation is trained using only token-alignment loss as described in the ablations here: https://huggingface.co/papers/2410.02678

Weights and Biases Run: https://wandb.ai/i18nlp/DiVA%20Training%20Runs/runs/4t0mvbcd?nw=nwuserheld

Citation

This is the token-alignment only model from https://huggingface.co/papers/2410.02678 BibTeX:

@misc{DiVA,
      title={{D}istilling an {E}nd-to-{E}nd {V}oice {A}ssistant {W}ithout {I}nstruction {T}raining {D}ata}, 
      author={William Held and Ella Li and Michael Ryan and Weiyan Shi and Yanzhe Zhang and Diyi Yang},
      year={2024},
      eprint={2410.02678},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.02678}, 
}
    

Table of Contents

• Model Card for DiVA Llama 3
• Citation
• Table of Contents
• Training Details
  • Training Data
  • Training Procedure
• Environmental Impact
• Technical Specifications [optional]
  • Model Architecture and Objective
  • Compute Infrastructure
    • Hardware
    • Software
• Model Card Contact

Training Details

Training Data

This model was trained on the CommonVoice corpus.

Training Procedure

This model was trained for 7k gradient steps with a batch size of 512 Recordings and a linearly decaying learning rate from 5e-5 to zero, with a linear warmup of 70 steps.

Environmental Impact

• Hardware Type: V4-32 TPU
• Hours used: 8 Hours
• Cloud Provider: Google Cloud.
• Compute Region: US Central C

Hardware

This model was trained on at V4 TPU on Google Cloud.

Software

This model was trained with Levanter

Model Card Authors [optional]

Will Held

Model Card Contact

held@stanford.edu