README.md

metadata

language:
  - en
license: apache-2.0
size_categories:
  - n<1K
task_categories:
  - text-generation
  - question-answering
  - text2text-generation
tags:
  - synthetic
  - instruction-tuning
  - bioinformatics
  - data-engineering
  - pandas
  - bash
  - aws
  - cron
  - json
  - qlora
pretty_name: Bio-DevOps Synthetic Instructions

Bio-DevOps Synthetic Instructions

This dataset contains synthetic instruction-following examples for biomedical-style data-engineering and scientific-computing workflows.

It was created for educational and portfolio use as part of a LoRA/QLoRA fine-tuning project using Qwen/Qwen2.5-Coder-7B-Instruct.

Related model:

AiLLMBS/qwen25-coder-bio-devops-lora

Dataset Contents

The dataset includes synthetic examples for:

• Python CSV validation
• pandas duplicate checks
• bash mount checks
• AWS S3 sync command generation
• cron expression explanation
• FASTQ manifest generation
• reproducible workflow checklist generation
• structured JSON extraction from synthetic workflow messages

Files

• train.jsonl
• eval.jsonl

Each row contains a chat-style instruction example with:

• messages
• meta.task_type
• meta.expected_keywords
• meta.requires_json

Data Source

This dataset is synthetically generated by project scripts.

It does not contain:

• PHI
• private employer data
• proprietary tickets
• internal emails
• client-specific workflows
• copyrighted book text

The examples use synthetic sample IDs, site IDs, file paths, S3 bucket names, and workflow messages.

Intended Use

This dataset is intended for:

• educational LoRA/QLoRA fine-tuning
• instruction-format experimentation
• synthetic code-assistant training examples
• portfolio demonstration of dataset creation and model evaluation
• safe local experimentation with coding-focused LLM adapters

Not Intended For

This dataset is not intended for:

• clinical decision support
• medical diagnosis
• training on PHI
• reproducing private employer workflows
• production automation without human review

Limitations

The dataset is synthetic and template-based. Models trained on it may overfit to repeated patterns and may not generalize well to diverse real-world workflows.

Generated code, shell commands, and AWS commands should always be reviewed before use.

Recommended Evaluation

Models trained on this dataset should be evaluated with:

• held-out prompts
• keyword coverage
• JSON validity checks
• unit tests for generated Python code
• manual review of shell commands
• base-model vs adapter comparison
• failure-case analysis

License

This dataset is released under the Apache-2.0 license.