rombodawg/Everyone-Coder-33b-Base

Everyone-Coder-33b-Base

EveryoneLLM series of models made by the community, for the community. This is a coding specific model made using fine-tunes of deekseekcoder-33b-base.

---

Im having trouble benchmarking this model because I suck at running llm benchmarks, but from hand testing running the model through https://edabit.com/challenge coding challenges vs up to date gpt-4. My model is hands down beating it in coding.

---

Ive recently noticed this model has trouble with end tokens so I made a custom prompt template for it. Made sure to add (Always end with "<|EOT|>") In addition to your system prompt and (Always end your response with "<|EOT|>") at the end of the User message is the preset. Then add <|EOT|> as a custom stop string in your LM text generating interface.

Always end with "<|EOT|>"

{System}

<|User|>

{User}. Always end your response with "<|EOT|>"

<|Assistant|>

{Assistant}

The models that were used in this merger were as follow:

• https://huggingface.co/deepseek-ai/deepseek-coder-33b-instruct

• https://huggingface.co/codefuse-ai/CodeFuse-DeepSeek-33B

• https://huggingface.co/WizardLM/WizardCoder-33B-V1.1

Thank you to the creators of the above ai models, they have full credit for the EveryoneLLM series of models. Without their hard work we wouldnt be able to achieve the great success we have in the open source community. 💗

You can find the write up for merging models here:

https://docs.google.com/document/d/1_vOftBnrk9NRk5h10UqrfJ5CDih9KBKL61yvrZtVWPE/edit?usp=sharing

Config for the merger can be found bellow:

models:
  - model: WizardLM_WizardCoder-33B-V1.1
    parameters:
      density: 1
      weight: .5
  - model: codefuse-ai_CodeFuse-DeepSeek-33B
    parameters:
      density: 1
      weight: .5
merge_method: ties
base_model: deepseek-ai_deepseek-coder-33b-instruct
parameters:
  normalize: true
  int8_mask: true
dtype: float16

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	49.48
AI2 Reasoning Challenge (25-Shot)	45.99
HellaSwag (10-Shot)	61.71
MMLU (5-Shot)	44.05
TruthfulQA (0-shot)	42.26
Winogrande (5-shot)	63.06
GSM8k (5-shot)	39.80