modular-ai

Code containing a modular AI architecture, using open-source models for various things. It includes, at the moment:

• Web crawling with Trafilatura
• Podcast generation with Llama

Step 0: Web Content Extraction (Optional)

If your source content is a webpage rather than a PDF or text file, start with this step:

For a single webpage:

python scripts/web_extractor.py https://example.com/article output.txt

Expected output:

• A JSON file containing the extracted content
• A clean text version of the webpage content
• Console output showing extraction progress and word count

For multiple webpages:

python scripts/web_extractor.py --file urls.txt output_directory/

Expected output:

• JSON files for each successfully processed URL
• An extraction summary file
• Console output showing progress for each URL

The extracted text can then be fed into Step 1 or directly into Step 2 depending on how clean the extraction is.

Note: This step is optional and should be used only if your source content is a webpage. For PDFs, start with Step 1.

NotebookLlama: An Open Source version of NotebookLM

Listen to audio from the example here

This is a guided series of tutorials/notebooks that can be taken as a reference or course to build a PDF to Podcast workflow.

You will also learn from the experiments of using Text to Speech Models.

It assumes zero knowledge of LLMs, prompting and audio models, everything is covered in their respective notebooks.

Outline:

Here is step by step thought (pun intended) for the task:

• Step 1: Pre-process PDF: Use Llama-3.2-1B-Instruct to pre-process the PDF and save it in a .txt file.
• Step 2: Transcript Writer: Use Llama-3.1-70B-Instruct model to write a podcast transcript from the text
• Step 3: Dramatic Re-Writer: Use Llama-3.1-8B-Instruct model to make the transcript more dramatic
• Step 4: Text-To-Speech Workflow: Use parler-tts/parler-tts-mini-v1 and bark/suno to generate a conversational podcast

Note 1: In Step 1, we prompt the 1B model to not modify the text or summarize it, strictly clean up extra characters or garbage characters that might get picked due to encoding from PDF. Please see the prompt in Notebook 1 for more details.

Note 2: For Step 2, you can also use Llama-3.1-8B-Instruct model, we recommend experimenting and trying if you see any differences. The 70B model was used here because it gave slightly more creative podcast transcripts for the tested examples.

Note 3: For Step 4, please try to extend the approach with other models. These models were chosen based on a sample prompt and worked best, newer models might sound better. Please see Notes for some of the sample tests.

Detailed steps on running the notebook:

Requirements: GPU server or an API provider for using 70B, 8B and 1B Llama models. For running the 70B model, you will need a GPU with aggregated memory around 140GB to infer in bfloat-16 precision.

Note: For our GPU Poor friends, you can also use the 8B and lower models for the entire pipeline. There is no strong recommendation. The pipeline below is what worked best on first few tests. You should try and see what works best for you!

• Before getting started, please make sure to login using the huggingface cli and then launch your jupyter notebook server to make sure you are able to download the Llama models.

You'll need your Hugging Face access token, which you can get at your Settings page here. Then run huggingface-cli login and copy and paste your Hugging Face access token to complete the login to make sure the scripts can download Hugging Face models if needed.

• First, please Install the requirements from here by running inside the folder:

git clone https://github.com/meta-llama/llama-recipes
cd llama-recipes/recipes/quickstart/NotebookLlama/
pip install -r requirements.txt

• Notebook 1:

This notebook is used for processing the PDF and processing it using the new Feather light model into a .txt file.

Update the first cell with a PDF link that you would like to use. Please decide on a PDF to use for Notebook 1, it can be any link but please remember to update the first cell of the notebook with the right link.

Please try changing the prompts for the Llama-3.2-1B-Instruct model and see if you can improve results.

• Notebook 2:

This notebook will take in the processed output from Notebook 1 and creatively convert it into a podcast transcript using the Llama-3.1-70B-Instruct model. If you are GPU rich, please feel free to test with the 405B model!

Please try experimenting with the System prompts for the model and see if you can improve the results and try the 8B model as well here to see if there is a huge difference!

• Notebook 3:

This notebook takes the transcript from earlier and prompts Llama-3.1-8B-Instruct to add more dramatization and interruptions in the conversations.

There is also a key factor here: we return a tuple of conversation which makes our lives easier later. Yes, studying Data Structures 101 was actually useful for once!

For our TTS logic, we use two different models that behave differently with certain prompts. So we prompt the model to add specifics for each speaker accordingly.

Please again try changing the system prompt and see if you can improve the results. We encourage testing the feather light 3B and 1B models as well at this stage

• Notebook 4:

Finally, we take the results from last notebook and convert them into a podcast. We use the parler-tts/parler-tts-mini-v1 and bark/suno models for a conversation.

The speakers and the prompt for parler model were decided based on experimentation and suggestions from the model authors. Please try experimenting, you can find more details in the resources section.

Note: Right now there is one issue: Parler needs transformers 4.43.3 or earlier and for steps 1 to 3 of the pipeline you need latest, so we just switch versions in the last notebook.

Next-Improvements/Further ideas:

• Speech Model experimentation: The TTS model is the limitation of how natural this will sound. This probably be improved with a better pipeline and with the help of someone more knowledgable-PRs are welcome! :)
• LLM vs LLM Debate: Another approach of writing the podcast would be having two agents debate the topic of interest and write the podcast outline. Right now we use a single LLM (70B) to write the podcast outline
• Testing 405B for writing the transcripts
• Better prompting
• Support for ingesting a website, audio file, YouTube links and more. Again, we welcome community PRs!

Resources for further learning:

• https://betterprogramming.pub/text-to-audio-generation-with-bark-clearly-explained-4ee300a3713a
• https://colab.research.google.com/drive/1dWWkZzvu7L9Bunq9zvD-W02RFUXoW-Pd?usp=sharing
• https://colab.research.google.com/drive/1eJfA2XUa-mXwdMy7DoYKVYHI1iTd9Vkt?usp=sharing#scrollTo=NyYQ--3YksJY
• https://replicate.com/suno-ai/bark?prediction=zh8j6yddxxrge0cjp9asgzd534
• https://suno-ai.notion.site/8b8e8749ed514b0cbf3f699013548683?v=bc67cff786b04b50b3ceb756fd05f68c

Step 0: Web Content Extraction (Optional)

If your source content is a webpage rather than a PDF or text file, start with this step:

For a single webpage:

python scripts/web_extractor.py https://example.com/article output.txt

Expected output:

• A JSON file containing the extracted content
• A clean text version of the webpage content
• Console output showing extraction progress and word count

For multiple webpages:

python scripts/web_extractor.py --file urls.txt output_directory/

Expected output:

• JSON files for each successfully processed URL
• An extraction summary file
• Console output showing progress for each URL

The extracted text can then be fed into Step 1 or directly into Step 2 depending on how clean the extraction is.

Note: This step is optional and should be used only if your source content is a webpage. For PDFs, start with Step 1.