Beating GPT-4 on HumanEval with a Fine-Tuned CodeLlama-34B

cross-posted from: https://lemmy.world/post/3879861

Beating GPT-4 on HumanEval with a Fine-Tuned CodeLlama-34B

Hello everyone! This post marks an exciting moment for [email protected] and everyone in the open-source large language model and AI community.

We appear to have a new contender on the block, a model apparently capable of surpassing OpenAI’s state of the art ChatGPT-4 in coding evals (evaluations).

This is huge. Not too long ago I made an offhand comment on us catching up to GPT-4 within a year. I did not expect that prediction to end up being reality in half the time. Let’s hope this isn’t a one-off scenario and that we see a new wave of open-source models that begin to challenge OpenAI.

Buckle up, it’s going to get interesting!

Here’s some notes from the blog, which you should visit and read in its entirety:

https://www.phind.com/blog/code-llama-beats-gpt4

Blog Post

We have fine-tuned CodeLlama-34B and CodeLlama-34B-Python on an internal Phind dataset that achieved 67.6% and 69.5% pass@1 on HumanEval, respectively. GPT-4 achieved 67% according to their official technical report in March. To ensure result validity, we applied OpenAI’s decontamination methodology to our dataset.

The CodeLlama models released yesterday demonstrate impressive performance on HumanEval.

CodeLlama-34B achieved 48.8% pass@1 on HumanEval

CodeLlama-34B-Python achieved 53.7% pass@1 on HumanEval

We have fine-tuned both models on a proprietary dataset of ~80k high-quality programming problems and solutions. Instead of code completion examples, this dataset features instruction-answer pairs, setting it apart structurally from HumanEval. We trained the Phind models over two epochs, for a total of ~160k examples. LoRA was not used — both models underwent a native fine-tuning. We employed DeepSpeed ZeRO 3 and Flash Attention 2 to train these models in three hours using 32 A100-80GB GPUs, with a sequence length of 4096 tokens.

Furthermore, we applied OpenAI’s decontamination methodology to our dataset to ensure valid results, and found no contaminated examples.

The methodology is:

For each evaluation example, we randomly sampled three substrings of 50 characters or used the entire example if it was fewer than 50 characters.

A match was identified if any sampled substring was a substring of the processed training example.

For further insights on the decontamination methodology, please refer to Appendix C of OpenAI’s technical report. Presented below are the pass@1 scores we achieved with our fine-tuned models:

Phind-CodeLlama-34B-v1 achieved 67.6% pass@1 on HumanEval

Phind-CodeLlama-34B-Python-v1 achieved 69.5% pass@1 on HumanEval

Download

We are releasing both models on Huggingface for verifiability and to bolster the open-source community. We welcome independent verification of results.

https://huggingface.co/Phind/Phind-CodeLlama-34B-v1

https://huggingface.co/Phind/Phind-CodeLlama-34B-Python-v1

If you get a chance to try either of these models out, let us know how it goes in the comments below!

If you found anything about this post interesting, consider subscribing to [email protected].

Cheers to the power of open-source! May we continue the fight for optimization, efficiency, and performance.