areen09/gprMax-ft_Mistral_v0.2

gprMaxAI-ft

This model is a fine-tuned version of TheBloke/Mistral-7B-Instruct-v0.2-GPTQ on the IraGia/gprMax_Train dataset.The fine-tuning process utilized QLoRA, an efficient low-rank adaptation method, to optimize the model for specific tasks.

Model description

The gprMaxAI-ft model was fine-tuned to enhance performance on tasks related to the gprMax domain, leveraging a pre-trained instruction-following model. The training process aimed to refine the model’s ability to understand and generate responses specific to the dataset it was trained on.

Intended Uses

Research and experimentation in low-rank adaptation of large language models.

Enhancing performance in gprMax-related tasks through fine-tuned responses.

Applications that require efficient fine-tuning of large-scale transformers on domain-specific data.

Limitations

The model is optimized for the training dataset and may not generalize well to unrelated domains.

It inherits biases and limitations from the base model (Mistral-7B-Instruct-v0.2-GPTQ).

Requires appropriate hardware for inference due to its size and complexity

Training and evaluation data

The model was fine-tuned on the IraGia/gprMax_Train dataset. This dataset was used to adapt the base model to perform better on tasks related to the gprMax domain.

Training procedure

1. The dataset was loaded using the datasets library.
2. Tokenization was performed using a customized tokenization function that truncates inputs to a maximum length of 512 tokens.
3. The dataset was split into training (80%) and validation (20%) sets.
4. The Mistral-7B-Instruct-v0.2-GPTQ model was loaded.
5. LoRA (Low-Rank Adaptation) was applied to enable efficient fine-tuning.
6. The optimizer used was paged_adamw_8bit with weight decay.
7. Training was performed with gradient checkpointing for memory efficiency.
8. The model was fine-tuned using the Trainer API with the defined arguments.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Use OptimizerNames.PAGED_ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 10
• mixed_precision_training: Native AMP

Framework versions

• PEFT 0.14.0
• Transformers 4.50.0
• Pytorch 2.6.0+cu124
• Datasets 3.5.0
• Tokenizers 0.21.1

Citation

If you use this model, consider citing the original base model and QLoRA methodology:

Mistral-7B-Instruct-v0.2-GPTQ

QLoRA: Efficient Finetuning of Quantized LLMs