This model is a fine-tuned version of codellama/CodeLlama-7b-hf on the AshtonIsNotHere/nlp_pp_code_dataset dataset. It achieves the following results on the evaluation set:

Loss: 0.4129
Accuracy: 0.8968

Model description

This model has been fine-tuned for code completion on a dataset of NLP++ code.

Intended uses & limitations

More information needed

Training and evaluation data

Dataset consists of a combination of scraped NLP++ code and NLP++ code examples from the VisualText website.

Training procedure

This model is trained in a multinode, multi-gpu setup with DeepSpeed Z3. For more information on the training setup, check out the GitHub repo.

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.00012
train_batch_size: 1
eval_batch_size: 1
seed: 42
distributed_type: multi-GPU
num_devices: 4
gradient_accumulation_steps: 4
total_train_batch_size: 16
total_eval_batch_size: 4
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_ratio: 0.1
num_epochs: 7.0

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
No log	1.0	61	0.5100	0.8726
No log	1.99	122	0.4129	0.8968
No log	2.99	183	0.4166	0.9072
No log	4.0	245	0.4595	0.9090
No log	5.0	306	0.5181	0.9093
No log	5.99	367	0.5553	0.9090
No log	6.97	427	0.5603	0.9089

Framework versions

Transformers 4.30.2
Pytorch 2.0.1+cu117
Datasets 2.13.0
Tokenizers 0.13.3

AshtonIsNotHere/CodeLlama_7B_nlp_pp

CodeLlama_7B_nlp_pp