nieche/Llama-2-7b-agriculture-niexche

Model Card for LLaMA-2-7B-NIEXCHE

This model was fine-tuned from LLaMA-2-7B on a Turkish agriculture QA dataset. It supports both Turkish and English languages and was trained for use in agriculture-related natural language processing (NLP) tasks.

Model Details

Model Description

• Developed by: NIEXCHE (Fevzi KILAS)
• Finetuned from model: meta-llama/Llama-2-7b
• License: Apache-2.0
• Language(s) (NLP): Turkish, English
• Model type: LLaMA-2-based model
• Training Dataset: NIEXCHE/turkish_agriculture_QA_llama2_22.6k

Model Sources

• Repository: Model Repository (TBA)
• Demo: TBA

Uses

Direct Use

The model can be used directly for question-answering tasks related to agriculture in Turkish and English. It is fine-tuned specifically for agricultural Q&A, making it suitable for similar domains and use cases.

Out-of-Scope Use

The model might not perform well on general knowledge questions outside of the agriculture domain.

Training Details

Training Data

The training data was a custom dataset created by translating and cleaning agricultural QA data from this source. The dataset contains 22.6k question-answer pairs in Turkish.

Training Procedure

The model was trained using the following frameworks and libraries:

• Frameworks: PyTorch, transformers, accelerate==0.21.0, peft==0.4.0, bitsandbytes==0.40.2, trl==0.4.7
• Precision: The model was trained using 4-bit quantization (BNB) with mixed precision (float16) to optimize memory usage.

Training Hyperparameters

• Base Model: meta-llama/Llama-2-7b
• Batch Size: 4 (per device)
• Learning Rate: 2e-4
• LoRA Parameters:
  • lora_r = 64
  • lora_alpha = 16
  • lora_dropout = 0.1
• Epochs: 1
• Optimizer: Paged AdamW (32-bit)
• Gradient Accumulation Steps: 1
• Scheduler: Cosine
• Max Gradient Norm: 0.3
• Gradient Checkpointing: Enabled
• Warmup Ratio: 0.03
• Group by Length: Enabled
• Max Sequence Length: None

Hardware

• Training Hardware: Google Colab Pro (A100 GPU) and 53 GB system RAM.
• Training Time: Approximately 1 hour 40 minutes.

Training output: TrainOutput(global_step=5654, training_loss=0.7829279924898043, metrics={'train_runtime': 6029.996, 'train_samples_per_second': 3.75, 'train_steps_per_second': 0.938, 'total_flos': 5.516196145999872e+16, 'train_loss': 0.7829279924898043, 'epoch': 1.0})

Evaluation

The same dataset (NIEXCHE/turkish_agriculture_QA_llama2_22.6k) was used for evaluation purposes.

Environmental Impact

Carbon emissions were estimated using the Machine Learning Impact calculator.

• Hardware Type: Google Colab (A100 GPU)
• Hours used: 1 hour 40 minutes
• Compute Region: Google Cloud (Colab)
• Carbon Emitted: Estimations pending

Citation

If you use this model in your research or applications, please cite it as:

@misc{Fevzi2024LLaMA-2-7B-NIEXCHE,
  author = {Fevzi KILAS},
  title = {LLaMA-2-7B-NIEXCHE: A Turkish Agriculture QA Model},
  year = {2024},
  howpublished = {https://huggingface.co./NIEXCHE/turkish_agriculture_QA_llama2_22.6k}
}

Contact:

NIEXCHE (Fevzi KILAS)