twinkle-ai/twinkle-sqlcoder

Devstral SQLCoder SFT

This model is a full-parameter SFT checkpoint for SQL generation, trained from mistralai/Devstral-Small-2505 and exported to Hugging Face safetensors format.

Model Details

• Base model: mistralai/Devstral-Small-2505
• Architecture: MistralForCausalLM
• Precision used in training: bf16
• Max sequence length (training config): 4096
• Export format: sharded safetensors with model.safetensors.index.json

Training Data (Merged)

The SFT run merged the following datasets:

• spider
• bird
• bird23-train-filtered
• synsql-2.5m
• wikisql
• gretelai-synthetic
• sql-create-context

Intended Use

• Text-to-SQL research and experimentation
• SQL generation benchmarks and evaluation pipelines

Limitations

• This model may generate incorrect SQL and should be validated before production use.
• Performance depends on prompt format, schema context quality, and decoding settings.
• Evaluate safety and compliance requirements before deployment.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_or_path = "<hf-username-or-org>/<model-repo>"

tokenizer = AutoTokenizer.from_pretrained(repo_or_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo_or_path,
    torch_dtype="bfloat16",
)

Local Files Included

• config.json
• generation_config.json
• tekken.json
• model-00001-of-00021.safetensors ... model-00021-of-00021.safetensors
• model.safetensors.index.json

Citation

If you use this model, please cite this repository:

• https://github.com/ai-twinkle/twinkle-sqlcoder