|
|
--- |
|
|
library_name: peft |
|
|
tags: |
|
|
- Llama2 |
|
|
- code |
|
|
- llama |
|
|
- opensource |
|
|
pipeline_tag: text-generation |
|
|
inference: true |
|
|
base_model: codellama/CodeLlama-34b-Instruct-hf |
|
|
--- |
|
|
|
|
|
## Aria Code is based on Code LLAMA 34B Instruct finetuned on French |
|
|
|
|
|
Aria code is a model built for educational purposes in French speaking countries on coding. |
|
|
|
|
|
The model is built to help students to learn coding in their native language while having the base qualities of LLAMA CODE which has been |
|
|
trained over 500 billions token of coding content. We belive coding skills are very valuable to reduce youth unemployment rates around the world |
|
|
and gives more scientific skills to students. LLAMA 2 base models have enough safeguards and censorship to ensure a safe use by kids and within academic environments. |
|
|
|
|
|
GPU used for training: Nvidia A100 |
|
|
|
|
|
Timing: Less than 24 hours |
|
|
|
|
|
Number of finetuning tokens : Over 10.000 high quality french language tokens. |
|
|
|
|
|
## Training procedure |
|
|
|
|
|
|
|
|
The following `bitsandbytes` quantization config was used during training: |
|
|
- quant_method: bitsandbytes |
|
|
- load_in_8bit: True |
|
|
- load_in_4bit: False |
|
|
- llm_int8_threshold: 6.0 |
|
|
- llm_int8_skip_modules: None |
|
|
- llm_int8_enable_fp32_cpu_offload: False |
|
|
- llm_int8_has_fp16_weight: False |
|
|
- bnb_4bit_quant_type: fp4 |
|
|
- bnb_4bit_use_double_quant: False |
|
|
- bnb_4bit_compute_dtype: float32 |
|
|
### Framework versions |
|
|
|
|
|
|
|
|
- PEFT 0.6.0.dev0 |
