Converted version of Mixtral 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.
Evaluation of the model was conducted using the PoLL (Pool of LLM) technique, assessing performance on 100 French questions with scores aggregated from six evaluations (two per evaluator). The evaluators included GPT-4o, Gemini-1.5-pro, and Claude3.5-sonnet.
Performance Scores (on a scale of 5):
| Model | Score | # params | size (GB) |
|---|---|---|---|
| gpt-4o | 4.13 | N/A | N/A |
| mistralai/Mixtral-8x7B-Instruct-v0.1 | 3.71 | 46.7b | 93.4 |
| cmarkea/Mixtral-8x7B-Instruct-v0.1-4bit | 3.68 | 46.7b | 23.35 |
| gpt-3.5-turbo | 3.66 | 175b | 350 |
| TheBloke/Mixtral-8x7B-Instruct-v0.1-GPTQ | 3.56 | 46.7b | 46.7 |
| mistralai/Mistral-7B-Instruct-v0.2 | 1.98 | 7.25b | 14.5 |
| cmarkea/bloomz-7b1-mt-sft-chat | 1.69 | 7.07b | 14.14 |
| cmarkea/bloomz-3b-dpo-chat | 1.68 | 3b | 6 |
| cmarkea/bloomz-3b-sft-chat | 1.51 | 3b | 6 |
| croissantllm/CroissantLLMChat-v0.1 | 1.19 | 1.3b | 2.7 |
| cmarkea/bloomz-560m-sft-chat | 1.04 | 0.56b | 1.12 |
| OpenLLM-France/Claire-Mistral-7B-0.1 | 0.38 | 7.25b | 14.5 |
The impact of quantization is negligible.
Here is a reminder of the command pattern to interact with the model if the add_special_tokens option is disabled (otherwise, do not include the BOS symbol and space at the beginning of the sequence):
<s> [INST] {user_prompt_1} [/INST] model_answer_1</s> [INST] {user_prompt_2} [/INST] model_answer_2</s>