code-millenials-34b / README.md
dittops's picture
Adding Evaluation Results (#2)
4fcd522 verified
metadata
license: llama2
library_name: transformers
tags:
  - code
model-index:
  - name: Code Millenials
    results:
      - task:
          type: text-generation
        dataset:
          name: HumanEval
          type: openai_humaneval
        metrics:
          - type: pass@1
            value: 0.8048
            name: pass@1
            verified: false
      - 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: 49.83
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=budecosystem/code-millenials-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: 75.09
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=budecosystem/code-millenials-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: 49.28
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=budecosystem/code-millenials-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: 45.37
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=budecosystem/code-millenials-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: 69.06
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=budecosystem/code-millenials-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: 32.45
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=budecosystem/code-millenials-34b
          name: Open LLM Leaderboard

Bud Code Millenials 34B

Welcome to our Code Model repository! Our model is specifically fine-tuned for code generation tasks. Bud Millenial Code Gen open-source models are currently the State of the Art (SOTA) for code generation, beating all the existing models of all sizes. We have achieved a HumanEval value of 80.48 @ Pass 1, beating proprietary models like Gemini Ultra, Claude, GPT-3.5 etc. by a large margin, and on par with GPT-4 (HumanEval ~ 82. Ref. WizardCoder). Our proprietary model (Bud Code Jr) beats GPT-4 as well with a HumanEval value of 88.2 & a context size of 168K, we will be releasing an API for Researchers, Enterprises, and potential Partners by January 2024 end. If interested, please reach out to [email protected]

News πŸ”₯πŸ”₯πŸ”₯

  • [2024/01/09] We released Code Millenials 3B , which achieves the 56.09 pass@1 on the HumanEval Benchmarks.
  • [2024/01/09] We released Code Millenials 1B , which achieves the 51.82 pass@1 on the HumanEval Benchmarks.
  • [2024/01/03] We released Code Millenials 34B , which achieves the 80.48 pass@1 on the HumanEval Benchmarks.
  • [2024/01/02] We released Code Millenials 13B , which achieves the 76.21 pass@1 on the HumanEval Benchmarks.

HumanEval

CodeMillenials

For the millenial models, the eval script in the github repo is used for the above result.

Note: The humaneval values of other models are taken from the official repos of WizardCoder, DeepseekCoder, Gemini etc.

Models

Model Checkpoint HumanEval (+) MBPP (+)
Code Millenials 34B HF Link 80.48 (75) 74.68 (62.9)
Code Millenials 13B HF Link 76.21 (69.5) 70.17 (57.6)
Code Millenials 3B HF Link 56.09 (52.43) 55.13 (47.11)
Code Millenials 1B HF Link 51.82 (48.17) 53.13 (44.61)

πŸš€ Quick Start

Inference code using the pre-trained model from the Hugging Face model hub

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("budecosystem/code-millenials-34b")
model = AutoModelForCausalLM.from_pretrained("budecosystem/code-millenials-34b")

template = """A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.

### Instruction: {instruction}

### Response:"""

instruction = <Your code instruction here>

prompt = template.format(instruction=instruction)

inputs = tokenizer(prompt, return_tensors="pt")
sample = model.generate(**inputs, max_length=128)
print(tokenizer.decode(sample[0]))

Training details

The model is trained of 16 A100 80GB for approximately 50hrs.

Hyperparameters Value
per_device_train_batch_size 16
gradient_accumulation_steps 1
epoch 3
steps 2157
learning_rate 2e-5
lr schedular type cosine
warmup ratio 0.1
optimizer adamw
fp16 True
GPU 16 A100 80GB

Important Note

  • Bias, Risks, and Limitations: Model may sometimes make errors, produce misleading contents, or struggle to manage tasks that are not related to coding.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric Value
Avg. 53.51
AI2 Reasoning Challenge (25-Shot) 49.83
HellaSwag (10-Shot) 75.09
MMLU (5-Shot) 49.28
TruthfulQA (0-shot) 45.37
Winogrande (5-shot) 69.06
GSM8k (5-shot) 32.45