Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
Data Science
Sebastian Raschka Blog·Jul 25·5 min read
Correction for Listing 6.5 in Build a Reasoning Model From Scratch
TL;DR
A correction was issued for the random seed in Listing 6.5 of Build a Reasoning Model From Scratch, ensuring reproducibility in experiments. This adjustment is crucial for accurate model training and evaluation.
✦ Why It Matters
Update your code with the correct random seed to ensure reproducibility in your experiments.
Key Takeaways
- Change torch.manual_seed(0) to torch.manual_seed(5) in your code.
- Ensure reproducibility of results in your experiments.
- Be aware of potential discrepancies in numeric outputs.
- Check for updates in future printings of the material.
Full Summary
In the blog 'Build a Reasoning Model From Scratch', a small typo was found in Listing 6.5, which affects reproducibility of results. The line that sets the random seed for PyTorch, torch.manual_seed(0), should actually be torch.manual_seed(5).
This change is crucial for obtaining the same generated response and log-probability outputs discussed later in Chapter 6. While using 0 does not break the code, it may lead to different numeric results, which could confuse users.
The author acknowledges this oversight and plans to correct it in the next printing of the material. This correction aims to save readers time during debugging and ensure they achieve the intended outcomes.