Public competition experiments
Parameter Golf
Study the tradeoffs behind a language model under a 16MB limit.
Public submissions and engineering notes from OpenAI's Parameter Golf challenge: model architecture, quantization, training kernels, and evaluation under strict artifact and compute limits.
What you can do with it
Who it's for: Engineers studying compact model training, quantization, and reproducible evaluation.
A concrete experiment to inspect or reproduce, with its code, logs, and submission discussion.
Your first useful result
- 01
Read one submission
Start with PR 1572 and review its architecture, run logs, artifact accounting, and reviewer discussion.
- 02
Choose an isolated technique
Compare one change, such as quantization or an attention kernel, against its stated baseline. Retain the upstream attribution and configuration when reproducing it.
- 03
Check the evidence state
Separate an author-reported run from an independently reproduced result or accepted leaderboard entry. Read the current PR status before citing a score.
Know the limits
- Public submissions include open and superseded PRs. Their reported scores do not by themselves establish an accepted record.
- Competition results are configuration- and date-specific; they are not a general language-model quality guarantee.