For the code to be accepted it must be measured and break the record. 🚀
Previously we were adding more features / code / architectures but records weren't getting broken and there was no progress. 📉
Important
Consider more code = bad (complexity, bloat, maintenance, bugs when upgrading), unless there is a new record in the training speed / loss, which justifies adding code. ✨
- 🔍 Pick a topic / task from issues (issues are general name for tasks), carefully read it and understand it
- 🍴 Fork the repo
- 💻 Clone it and implement the experiment, follow README
- 📊 Benchmark your changes against a the baseline that you also measured beforehand. If hardware is limited, use free GPUs (Lightning AI, Colab) and reduce model size (
n_layer,n_embd) for testing. - 📥 Submit a PR with your findings and comparison data.
Please check LEADERBOARD for architecture, records and contribution history.