The dataset is available for academic use, please see https://rowanzellers.com/merlot for access information. We do not plan to release the videos, just their metadata (like the video IDs, titles, and descriptions).
process.py contains a quick script for turning the example_video into a very small tfrecord for pretraining. You'll need to download videos into a similar format (including the Grover denoisified transcript), or change the script, to get things to work.
We've uploaded denoised transcripts as part of the dataset release. If you'd like the denoiser model we used, see the data/groverdenoise folder. Our checkpoint is publicly available at gs://merlot/denoisify_ckpt/model=medium~lr=1e-5~steps=80000~v0/model.ckpt-80000\* -- you'll need to download that to your computer, and run data/groverdenoise/run_server.py, which should let you interact with the model.