I don’t remember what it’s like to think with more than 60 seconds of context.

Recently, GDM released a great paper titled, Scaling Exponents Across Parameterizations and Optimizers, in which they conduct over 10,000 LLM training runs to obtain optimal hyperparameters under different regimes.

After reading it (it was great), I wanted to test my understanding of the paper by tallying up all experiments conducted within, calculating the total compute cost it would take to replicate the paper.

Paper summary: 1.3B to 33B LLMs on 1/2T code tokens (87 langs) w/ FiM and 16K seqlen. Strong effort in constructing pretraining data from Github from scratch, with repository-level samples. Evals beat OSS code models solidly + GPT-3.5 a bit; Coder-7B > CodeLlama-33B often.

They don’t spend much effort on Instruction tuning. They commit a continued pretrain of DeepSeek LLM -> Coder: I believe it underperforms; they don’t.

Paper summary: LLaMA-like 7B/67B pretrain (Base) + SFT&DPO (Chat). 2T tokens with strong CN/EN mix, >1mil SFT examples. Well-executed exploration of scaling laws. Good details about evals and safety. Not much described about their actual data.

I don’t think very often. These are some of the things I’ve tried in the past to rectify that.