Mistral-7B-text-to-sql-flash-attention-2-dataeval

This model is a fine-tuned version of mistralai/Mistral-7B-Instruct-v0.3 on the generator dataset. It achieves the following results on the evaluation set:

• Loss: 0.4605

Perplexity of 10.40

Perplexity: Perplexity is a measure of how uncertain or surprised the model is about its predictions. It's derived from the probabilities the model assigns to different words or tokens.

Perplexity Article: https://www.jmlr.org/papers/volume3/bengio03a/bengio03a.pdf https://medium.com/@AyushmanPranav/perplexity-calculation-in-nlp-0699fbda4594

The perplexity of 10.40 achieved on the dataset indicates that the fine-tuned Mistral-7B model reasonably understands natural language and SQL syntax. However, further evaluation using task-specific metrics is necessary to assess the model's effectiveness in real-world scenarios. By combining quantitative metrics like perplexity with qualitative analysis of generated queries, we can comprehensively understand the model's strengths and weaknesses, ultimately leading to improved performance and more reliable text-to-SQL translation capabilities.

Dataset : b-mc2/sql-create-context

Model description

Article: https://medium.com/@frankmorales_91352/fine-tuning-the-llm-mistral-7b-instruct-v0-3-249c1814ceaf

Training and evaluation data

Fine Tuning and Evaluation: https://github.com/frank-morales2020/MLxDL/blob/main/FineTuning_LLM_Mistral_7B_Instruct_v0_1_for_text_to_SQL_EVALDATA.ipynb

Evaluation: https://github.com/frank-morales2020/MLxDL/blob/main/Evaluator_Mistral_7B_text_to_sql.ipynb

Evaluation article with Chromadb: https://medium.com/@frankmorales_91352/a-comprehensive-evaluation-of-a-fine-tuned-text-to-sql-model-from-code-to-results-with-7ea59943b0a1

Evaluation article with Chromadb, PostgreSQL and the “gretelai/synthetic_text_to_sql” dataset: https://medium.com/@frankmorales_91352/evaluating-the-performance-of-a-fine-tuned-text-to-sql-model-6b7d61dcfef5 The article discusses evaluating this fine-tuned text-to-SQL model, a type of artificial intelligence that translates natural language into SQL queries.

The model was trained on the "b-mc2/sql-create-context" dataset and evaluated using the "gretelai/synthetic_text_to_sql" dataset.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 3
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 24
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant
• lr_scheduler_warmup_ratio: 0.03
• lr_scheduler_warmup_steps: 15
• num_epochs: 3

from transformers import TrainingArguments

args = TrainingArguments( output_dir="Mistral-7B-text-to-sql-flash-attention-2-dataeval",

num_train_epochs=3,                     # number of training epochs
per_device_train_batch_size=3,          # batch size per device during training
gradient_accumulation_steps=8,      #2  # number of steps before performing a backward/update pass
gradient_checkpointing=True,            # use gradient checkpointing to save memory
optim="adamw_torch_fused",              # use fused adamw optimizer
logging_steps=10,                       # log every ten steps
#save_strategy="epoch",                  # save checkpoint every epoch
learning_rate=2e-4,                     # learning rate, based on QLoRA paper
bf16=True,                              # use bfloat16 precision
tf32=True,                              # use tf32 precision
max_grad_norm=0.3,                      # max gradient norm based on QLoRA paper
warmup_ratio=0.03,                      # warmup ratio based on QLoRA paper
weight_decay=0.01,
lr_scheduler_type="constant",           # use constant learning rate scheduler
push_to_hub=True,                       # push model to hub
report_to="tensorboard",                # report metrics to tensorboard
hub_token=access_token_write,           # Add this line
load_best_model_at_end=True,
logging_dir="/content/drive/MyDrive/model/Mistral-7B-text-to-sql-flash-attention-2-dataeval/logs",
evaluation_strategy="steps",
eval_steps=10,
save_strategy="steps",
save_steps=10,
metric_for_best_model = "loss",
warmup_steps=15,

)

Training results

Training Loss	Epoch	Step	Validation Loss
1.8612	0.4020	10	0.6092
0.5849	0.8040	20	0.5307
0.4937	1.2060	30	0.4887
0.4454	1.6080	40	0.4670
0.425	2.0101	50	0.4544
0.3498	2.4121	60	0.4717
0.3439	2.8141	70	0.4605

Framework versions

• PEFT 0.11.1
• Transformers 4.41.2
• Pytorch 2.3.0+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1