This is an instruction fine-tuned model focussed on the task of patching code. Patching may include fixing bugs, remediating security vulnerabilities,
doing API migrations and other kinds of code maintenance.
This model has undergone very limited testing. Additional safety testing should be performed before any real-world deployments.
Training Details
GPU: A100 80 GB
Time: ~8 hrs
Training Data
The model was fine-tuned on commitpackft, an open dataset consisting of commits.
We started with the commits for the python langauge from the dataset and then filtered all the commits that were related to fixing bugs.
Training Procedure
Instruction fine-tuning to follow instructions in natural langauge related to code. We load the quantized base model in 4 bits
and then use QLoRA for Parameter-Efficient Fine-Tuning (PEFT) with Flash Attention. The model was trained for 2 epochs.
Training Hyperparameters
Training regime:
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: True
bnb_4bit_compute_dtype: bfloat16
Evaluation
We evaluated the model on HumanEval (for code generation) and HumanEvalFix Python (for bug fixing) benchmarks using
Code Generation LM Evaluation Harness.
To evaluate the model for vulnerability remediation we used the Static Analysis Eval benchmark available here.
Results
Model
HumanEval
HumanEval Fix Python
Static Analysis Eval
patched-coder-34b
53.57
41.34
51.32
CodeLlama-34b-Python
53.29
33.14
27.63
GPT-4
86.6
47
55.26
Based on the results on these benchmarks, patched-coder-34b is the SOTA open code LLM. Other code LLMs (e.g. from WizardCoder and Phind) are trained on
either unknown proprietary datasets or used OpenAI's APIs for training, thus making them unviable for commercial use.
