codellama2-finetuned-nl2bash-fin

Finetuned on the AnishJoshi/nl2bash-custom dataset for generating bash code based on natural language descriptions.

Model Details

• Model Name: CodeLlama2-Finetuned-NL2Bash
• Base Model: CodeLlama2
• Task: Natural Language to Bash Script Conversion
• Framework: PyTorch
• Fine-tuning Dataset: Custom dataset of natural language commands and corresponding Bash scripts, available here

Model Description

• Developed by: Anish Joshi
• Model type: CausalLM
• Finetuned from model: Codellama2

Files Included

• adapter_config.json: Configuration file for the adapter layers.
• adapter_model.safetensors: Weights of the adapter layers in the Safetensors format.
• optimizer.pt: State of the optimizer used during training.
• rng_state.pth: State of the random number generator.
• scheduler.pt: State of the learning rate scheduler.
• special_tokens_map.json: Mapping for special tokens used by the tokenizer.
• tokenizer.json: Tokenizer model including the vocabulary.
• tokenizer_config.json: Configuration settings for the tokenizer.
• trainer_state.json: State of the trainer including training metrics.
• training_args.bin: Training arguments used for fine-tuning.
• `README.md

Usage

Load model directly

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("AnishJoshi/codellama2-finetuned-nl2bash-fin")
model = AutoModelForCausalLM.from_pretrained("AnishJoshi/codellama2-finetuned-nl2bash-fin")

Training Details

Training details available at Finetuning Notebook

Training Hyperparameters

Training arguments and configuration are set using TrainingArguments and LoraConfig. The model is fine-tuned using the following parameters:

• output_dir: codellama2-finetuned-nl2bash - Directory to save the fine-tuned model.
• per_device_train_batch_size: 2 - Batch size per device.
• gradient_accumulation_steps: 16 - Number of gradient accumulation steps.
• optim: paged_adamw_32bit - Optimizer type.
• learning_rate: 2e-4 - Learning rate.
• lr_scheduler_type: cosine - Learning rate scheduler type.
• save_strategy: epoch - Strategy to save checkpoints.
• logging_steps: 10 - Number of steps between logging.
• num_train_epochs: 1 - Number of training epochs.
• max_steps: 100 - Maximum number of training steps.
• fp16: True - Use 16-bit floating-point precision.
• push_to_hub: False - Whether to push the model to Hugging Face Hub.
• report_to: none - Reporting destination.

Evaluation

Evaulation metrics and calculations available at Evaluation Notebook