Fine-Tune C++ JUCE Audio DSP Model, Dataset Curation & QLoRA

About OpenTrain

OpenTrain is the #1 platform for people starting and building careers in AI training and data labeling. We connect experienced contributors with hands-on projects that shape how modern AI systems behave.

OpenTrain AI is the hiring and contracting organization for this role — we run the project, provide resources, and hire contractors directly.

About AI training and why this work matters

AI training (data labeling / annotation / human feedback) is the human side of building AI: people create, curate, and review the examples models learn from. This work is often remote, flexible, and accessible, and contributors directly influence model behavior.

This project focuses on coding datasets and model fine-tuning for developer-facing code models — a high-impact area where precise examples and real-world code matter for model quality.

Role overview

We are hiring an expert ML engineer with strong JUCE/C++ audio DSP knowledge to build a supervised fine-tuning (SFT) dataset and run QLoRA training on Qwen3-Coder. The engagement is contract, part-time (~20+ hours/week), fixed-price.

You will extract DSP-relevant C++ functions from many open-source codebases, convert tutorial and reference material into ChatML instruction–response examples, apply LLM-assisted generation and quality filtering, deduplicate, and execute QLoRA fine-tuning using Unsloth.

What you'll do (daily tasks)

This is hands-on data curation and fine-tuning work that mixes coding, writing training examples, and running experiments. Expect a mix of code extraction, prompt engineering, LLM orchestration, and training runs.

• Extract DSP-relevant C++ functions and code snippets from 40+ open-source GitHub repositories (examples: Surge, ChowDSP, Airwindows, Vital, JUCE framework).
• Generate high-quality instruction–response pairs using LLM-assisted pipelines (Bespoke Curator or Distilabel with Claude/GPT-4) and human review.
• Convert blog posts, tutorials, forum Q&A, and free textbook content into clean ChatML-formatted training examples with clear instructions and model responses.
• Run a second LLM pass for quality filtering and perform deduplication and normalization of examples.
• Execute QLoRA fine-tuning on Qwen3-Coder using Unsloth, iterate on training settings, and validate outputs against held-out examples and DSP correctness checks.

Deliverables and targets

Deliver a high-quality SFT dataset and evidence of the fine-tuning run. Be precise: quality matters more than raw volume, but we have concrete targets to meet.

• Target dataset size: 3,000–5,000 instruction–response examples focused on processBlock, AudioBuffer, juce_dsp filters, oscillators, delay lines, reverb, virtual analog modeling, plugin architecture, and real-time DSP best practices.
• A cleaned ChatML training corpus, fully deduplicated and filtered (with notes on filtering criteria).
• A reproducible training log and scripts showing QLoRA runs on Qwen3-Coder using Unsloth, plus checkpoints and evaluation notes.
• A comprehensive resource document listing all repo URLs, blog links, textbook references, and the provided clone script (the resource doc and clone script will be provided to the hired candidate).

Requirements and preferences

You must be an expert in C++ audio plugin development and digital signal processing, with practical experience in JUCE. This is a specialist role — real examples matter.

• Essential: Expert-level knowledge of C++ audio plugin development (JUCE framework) and DSP concepts relevant to audio (filters, oscillators, delay, reverb, virtual analog modeling, real-time constraints).
• Essential: Demonstrated experience building datasets or SFT examples for code models, or experience running LLM-assisted data pipelines.
• Required: Ability to run QLoRA-style fine-tuning workflows (experience with QLoRA and Unsloth a strong plus).
• Preferred: Prior use of Bespoke Curator or Distilabel workflows with Claude/GPT-4 for instruction–response generation.
• Nice to have: Shareable JUCE/audio project examples or links to public repos demonstrating your work — seeing example audio projects is helpful.

Time, pay, and contract

This is a contract, part-time engagement with a fixed-price payment. Work is remote and open worldwide.

The project requires roughly 20+ hours/week; proposed schedule and milestones will be agreed at onboarding.

• Payment: Fixed price of USD 1,000 for the project as specified in the posting.
• Employment type: Contractor, part-time. Worldwide applicants welcome.
• Tools & access: The project will provide a comprehensive resource document and a clone script to access the target repositories and reference materials.

How to apply and next steps

Apply through OpenTrain and include a brief summary of relevant experience, links to any JUCE or audio DSP code samples, and a short plan outlining how you would extract code, create ChatML examples, and run QLoRA with Unsloth.

We will review applicants for technical fit and examples of prior JUCE/DSP work; successful candidates will receive the resource document and clone script and begin with a short onboarding milestone.

• When applying, highlight: JUCE experience, examples of audio projects (links preferred), past dataset or fine-tuning work, and familiarity with QLoRA/Unsloth if any.
• OpenTrain is the hiring organization — all contracting, resources, and payments are managed through OpenTrain.