ARIA CODE FR Instruct is a finetuned model based on LLAMA CODE 34B INSTRUCT
This model has been trained over 10 Millions tokens on a french language dataset structured with Alpaca style.
GPU used for training : NVIDIA A100
Timing of training: 24H
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
