Reproducing a Meta AI paper on free GPUs
I reproduced Meta's COCONUT paper on free GPUs. I got 95.33 percent against 97.0.
COCONUT is a Meta paper about letting a language model reason in its own hidden space instead of in written steps. I reproduced it on the ProsQA reasoning test, with GPT-2 as the base model.
I had no GPU of my own, so I trained on Kaggle's free GPUs. That took about 99.7 GPU-hours over nine weekly quota resets, from June to August 2026.
My COCONUT model scored 95.33 percent against the paper's 97.0. My written-steps baseline scored 80.0 percent against the paper's 77.5. That is close and still short, and I ran only one seed.
Things went wrong along the way. A checkpoint download came through cut short, one chunk trained from scratch instead of resuming, and an evaluation bug would have given wrong accuracy. Each one was caught and written up.
- PaperMeta FAIR, COCONUT (arXiv 2412.06769)
- Test setProsQA, 300 questions
- COCONUT95.33 percent (paper 97.0)
- Written-steps baseline80.0 percent (paper 77.5)
- ComputeKaggle free GPUs, about 99.7 GPU-hours
