Review

A technical report of InternLM on 6/7. It consisted of 104 billion parameters, was trained on 1.6 trillion tokens, and was fine-tuned for performance in Chinese.

The authors claimed that it performed second-best on the Chinese language benchmark C-Eval, right after GPT4. In addition, it performed at the level of GPT3.5 in one-shot MMLU. A version fine-tuned for programming also performed similarly to GPT3.5 in coding benchmarks like HumanEval.

Notable takeaways: 

  1. Significant effort was put into parallelization to help evade the US chip ban. I don't know how impressive this actually is.
  2. It achieved GPT3.5-level performance with similar-ish levels of compute and data. The China-America algorithmic gap is shrinking.
  3. My gut feeling is that the model was very specifically fine-tuned for performing well on standardized tests, especially those in Chinese (GK/Gao Kao is the Chinese College entrance exam). It was also consistently bad with math.
New Comment
4 comments, sorted by Click to highlight new comments since:

I’m confused about the parallelization part and what it implies. It says the model was trained on 2K GPUs but GPT-4 was probably trained on 1 OOM more than that right?

Parallelization part (data parallelism, tensor parallelism, pipeline parallelism, ZeRO) is completely standard. See Efficient Training on Multiple GPUs by Hugging Face for a standard description. Failure recovery part is relatively unusual.

I don't get what the parallelization strategy should have to do with the chip ban? It sounds like just a basic parallelism approach.

You're right. I was pretty tired when I wrote this and am not sure where that thought came from.