Model Description

These are the peft layers of the finetuned codellama 7b model that is to be used as the retriever component of RAG (Retrieval Augmented Generation) for codebases

• Developed by: arcee.ai
• Model type: Retriever component for RAG for codebases
• Language(s) (NLP): English
• License: apache-2.0
• Finetuned from model: codellama/CodeLlama-7b-Instruct-hf

Training Details

• Trained on a small dataset of size: 160,000 query-passage pairs
• The pretrained model codellama/CodeLlama-7b-Instruct-hf was finetuned using data containing code
• PEFT and quantization was used
• the hidden state associated with the EOS token was used as the embeddings
• Model was trained for 1 epoch

Training procedure

The following bitsandbytes quantization config was used during training:

• quant_method: bitsandbytes
• load_in_8bit: False
• load_in_4bit: True
• 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: nf4
• bnb_4bit_use_double_quant: False
• bnb_4bit_compute_dtype: bfloat16

Framework versions

• PEFT 0.5.0