Llama-3-8B-Instruct-MoE-2-spider

This model is a fine-tuned version of VictorDCh/Llama-3-8B-Instruct-MoE-2 on the generator dataset.

Model description

This is a mixture of experts merged with mergekit from 4 experts (Llama 3 8B Instruct), fine-tuned on the Spider train dataset.

Intended uses & limitations

This is part of a research project on using open-source models for text-to-SQL use cases.

Some pre-processing of the input data, and post-processing of the results may be necessary to obtain optimal results.

The input prompt uses the following system-user-assistant format:

• system message: """You are a text to SQL query translator. Users will ask you questions in English and you will generate a SQL query based on the provided SCHEMA. Only generate SQL. SCHEMA: {schema}"""
• user message: """Here is the user question in natural language: {question} """
• assistant message (empty when testing): {SQL query}

Training and evaluation data

The training and testing dataset can be found here: Cleaned Spider dataset.

Training procedure

One epoch of SFT on the train dataset as explained above.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 1
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 2
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 1

Framework versions

• PEFT 0.7.2.dev0
• Transformers 4.36.2
• Pytorch 2.1.2+cu121
• Datasets 2.16.1
• Tokenizers 0.15.2