| ---
|
| language:
|
| - en
|
| license: apache-2.0
|
| base_model: Qwen/Qwen2.5-Coder-7B-Instruct
|
| tags:
|
| - qwen
|
| - code
|
| - lora
|
| - qlora
|
| - bioinformatics
|
| - data-engineering
|
| - aws
|
| - pandas
|
| - bash
|
| library_name: peft
|
| ---
|
|
|
| # Qwen2.5-Coder Bioinformatics/Data-Engineering LoRA
|
|
|
| This repository contains a LoRA adapter fine-tuned from `Qwen/Qwen2.5-Coder-7B-Instruct` on a synthetic biomedical data-engineering instruction dataset.
|
|
|
| ## Intended Use
|
|
|
| This adapter is intended for educational and portfolio use. It is designed to help with tasks such as:
|
|
|
| - Python CSV validation
|
| - pandas data cleaning
|
| - bash scripting
|
| - AWS S3 command generation
|
| - cron explanation
|
| - FASTQ manifest generation
|
| - reproducible workflow checklists
|
|
|
| ## Not Intended For
|
|
|
| This model is not intended for:
|
|
|
| - clinical decision support
|
| - medical diagnosis
|
| - handling PHI
|
| - production automation without human review
|
| - private/proprietary workflow reproduction
|
|
|
| ## Base Model
|
|
|
| Base model: `Qwen/Qwen2.5-Coder-7B-Instruct`
|
|
|
| ## Training Method
|
|
|
| - Method: QLoRA / LoRA adapter training
|
| - Quantization: 4-bit NF4
|
| - LoRA rank: 16
|
| - LoRA alpha: 32
|
| - LoRA dropout: 0.05
|
| - Hardware: local NVIDIA RTX 5090 32 GB
|
| - Dataset: synthetic instruction examples generated by project scripts
|
|
|
| ## Dataset
|
|
|
| Training dataset:
|
|
|
| AiLLMBS/bio-devops-synthetic-instructions
|
|
|
| The dataset is synthetic and was generated for educational LoRA/QLoRA fine-tuning. It does not contain PHI, private employer data, proprietary tickets, internal emails, client-specific workflows, or copyrighted book text.
|
|
|
| The dataset includes synthetic examples for Python CSV validation, pandas duplicate checks, bash mount checks, AWS S3 command generation, cron expression explanation, FASTQ manifest generation, reproducible workflow checklist generation, and structured JSON extraction from synthetic workflow messages.
|
|
|
| ## Evaluation
|
|
|
| Final eval loss: `8.545802302251104e-06`
|
|
|
| Final eval perplexity: `1.0000085458388177`
|
|
|
| Average keyword coverage: `0.945`
|
|
|
| JSON validity rate: `1.0`
|
|
|
| These metrics are early portfolio metrics, not a claim of production quality. Generated code should be manually reviewed.
|
|
|
| ## Example Prompt
|
|
|
| ```text
|
| Write a Python script that validates a sample manifest CSV. It should check for required columns sample_id, site_id, and file_path, then report missing values and duplicate sample IDs.
|
| ```
|
|
|
| ## Limitations
|
|
|
| This adapter may hallucinate commands, produce code with bugs, or omit safety checks. All generated code should be reviewed before use.
|
|
|
