XwinCoder
We are glad to introduce our instruction finetuned code generation models based on CodeLLaMA: XwinCoder. We release model weights and evaluation code.
Repository: https://github.com/Xwin-LM/Xwin-LM/tree/main/Xwin-Coder
Models:
| Model | 🤗hf link | HumanEval pass@1 | MBPP pass@1 | APPS-intro pass@5 |
|---|---|---|---|---|
| XwinCoder-7B | link | 63.8 | 57.4 | 31.5 |
| XwinCoder-13B | link | 68.8 | 60.1 | 35.4 |
| XwinCoder-34B | link | 74.2 | 64.8 | 43.0 |
Updates
-
💥 We released XwinCoder-7B, XwinCoder-13B, XwinCoder-34B. Our XwinCoder-34B reached 74.2 on HumanEval and it achieves comparable performance as GPT-3.5-turbo on 6 benchmarks.
-
We support evaluating instruction finetuned models on HumanEval, MBPP, APPS, DS1000 and MT-Bench. See our github repository.
Overview
- To fully demonstrate our model's coding capabilities in real-world usage scenarios, we have conducted thorough evaluations on several existing mainstream coding capability leaderboards (rather than only on the currently most popular HumanEval).
- As shown in the radar chart results, our 34B model achieves comparable performance as GPT-3.5-turbo on coding abilities.
- It is worth mentioning that, to ensure accurate visualization, our radar chart has not been scaled (only translated; MT-Bench score is scaled by 10x to be more comparable with other benchmarks).
- Multiple-E-avg6 refer to the 6 languages used in CodeLLaMA paper. Results of GPT-4 and GPT-3.5-turbo are conducted by us, more details will be released later.
Demo
We provide a chat demo in our github repository, here are some examples:
