Converted version of Qwen2.5-Coder-7B-Instruct to 4-bit using bitsandbytes. For more information about the model, refer to the model's page.
Impact of quantization on a set of models.
We evaluated the models using the PoLL (Pool of LLM) technique a panel of giga-models (GPT-4o, Gemini Pro 1.5, and Claude-Sonnet 3.5). The scoring ranged from 0, indicating a model unsuitable for the task, to 5, representing a model that fully met expectations. The evaluation was based on 67 instructions across four programming languages: Python, Java, JavaScript, and Pseudo-code. All tests were conducted in a French-language context, and models were heavily penalized if they responded in another language, even if the response was technically correct.
Performance Scores (on a scale of 5):
| Model | Score | # params (Billion) | size (GB) |
|---|---|---|---|
| gemini-1.5-pro | 4.51 | NA | NA |
| gpt-4o | 4.51 | NA | NA |
| claude3.5-sonnet | 4.49 | NA | NA |
| deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct | 4.24 | 15.7 | 31.4 |
| meta-llama/Meta-Llama-3.1-70B-Instruct | 4.23 | 70.06 | 141.2 |
| cmarkea/Meta-Llama-3.1-70B-Instruct-4bit | 4.14 | 70.06 | 35.3 |
| Qwen/Qwen2.5-Coder-7B-Instruct | 4.11 | 7.62 | 15.24 |
| cmarkea/Qwen2.5-Coder-7B-Instruct-4bit | 4.08 | 7.62 | 3.81 |
| cmarkea/Mixtral-8x7B-Instruct-v0.1-4bit | 3.8 | 46.7 | 23.35 |
| meta-llama/Meta-Llama-3.1-8B-Instruct | 3.73 | 8.03 | 16.06 |
| mistralai/Mixtral-8x7B-Instruct-v0.1 | 3.33 | 46.7 | 93.4 |
| codellama/CodeLlama-13b-Instruct-hf | 3.33 | 13 | 26 |
| codellama/CodeLlama-34b-Instruct-hf | 3.27 | 33.7 | 67.4 |
| codellama/CodeLlama-7b-Instruct-hf | 3.19 | 6.74 | 13.48 |
| cmarkea/CodeLlama-34b-Instruct-hf-4bit | 3.12 | 33.7 | 16.35 |
| codellama/CodeLlama-70b-Instruct-hf | 1.82 | 69 | 138 |
| cmarkea/CodeLlama-70b-Instruct-hf-4bit | 1.64 | 69 | 34.5 |
The impact of quantization is negligible.
Here is a reminder of the command pattern to interact with the model:
<|im_start|>user\n{user_prompt_1}<|im_end|><|im_start|>assistant\n{model_answer_1}<|im_end|>...