metadata

license: llama2
model-index:
  - name: XwinCoder-34B
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: AI2 Reasoning Challenge (25-Shot)
          type: ai2_arc
          config: ARC-Challenge
          split: test
          args:
            num_few_shot: 25
        metrics:
          - type: acc_norm
            value: 51.02
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Xwin-LM/XwinCoder-34B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: HellaSwag (10-Shot)
          type: hellaswag
          split: validation
          args:
            num_few_shot: 10
        metrics:
          - type: acc_norm
            value: 74.02
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Xwin-LM/XwinCoder-34B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU (5-Shot)
          type: cais/mmlu
          config: all
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 49.53
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Xwin-LM/XwinCoder-34B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: TruthfulQA (0-shot)
          type: truthful_qa
          config: multiple_choice
          split: validation
          args:
            num_few_shot: 0
        metrics:
          - type: mc2
            value: 43.82
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Xwin-LM/XwinCoder-34B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: Winogrande (5-shot)
          type: winogrande
          config: winogrande_xl
          split: validation
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 68.35
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Xwin-LM/XwinCoder-34B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GSM8k (5-shot)
          type: gsm8k
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 39.35
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Xwin-LM/XwinCoder-34B
          name: Open LLM Leaderboard

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:

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	54.35
AI2 Reasoning Challenge (25-Shot)	51.02
HellaSwag (10-Shot)	74.02
MMLU (5-Shot)	49.53
TruthfulQA (0-shot)	43.82
Winogrande (5-shot)	68.35
GSM8k (5-shot)	39.35