Posted: Feb 27, 2026
Fine-Tune Qwen3-Coder on JUCE/C++ Audio DSP — Dataset Curation & QLoRA Training
Total Budget: $1,000OtherComputer Code ProgrammingFixed Price
Overview
Dataset
Labeling Details
Hiring
Budget
Client
Labeling Overview
Labeling Overview: We need an ML engineer to build a supervised fine-tuning (SFT) dataset for a JUCE/C++ audio DSP coding model. The work involves: (1) extracting DSP-relevant C++ functions from 40+ open-source GitHub repos (Surge, ChowDSP, Airwindows, Vital, JUCE framework, etc.), (2) generating high-quality instruction-response pairs using LLM-assisted pipelines (Bespoke Curator or Distilabel with Claude/GPT-4), (3) converting blog posts, tutorials, forum Q&A, and free textbook content into clean ChatML-formatted training examples, (4) quality filtering with a second LLM pass and deduplication, and (5) running QLoRA fine-tuning on Qwen3-Coder using Unsloth. Target: 3,000–5,000 examples covering processBlock, AudioBuffer, juce_dsp filters, oscillators, delay lines, reverb, virtual analog modeling, plugin architecture, and real-time DSP best practices. A comprehensive resource document with all repo URLs, blog links, textbook references, and a clone script will be provided to the hired candidate.
Total Budget: $1,000Fixed PriceExpert1-3 monthsAgency or Individual
Total Budget
$1,000
Pay per Label
-
Time Requirement
20+ hrs/week
Duration
1-3 months
Description of dataset
3,000–5,000 instruction-response pairs for fine-tuning a code LLM on JUCE C++ audio plugin development and DSP (digital signal processing). Sourced from 40+ open-source GitHub repos, free DSP textbooks, technical blogs, official JUCE tutorials/API docs, and forum Q&A. Includes function-level code extraction, natural-language explanations, and synthetic educational examples covering topics like processBlock, AudioBuffer, filters, oscillators, reverb, virtual analog modeling, and plugin architecture.
Software
Other
Hiring Type
Agency or Individual
Required Location
Global Any Location
Workload / Schedule
This is a 2–4 week core project with flexibility for iteration. Phase 1 (dataset curation) is the most time-intensive at roughly 40–60 hours, followed by Phase 2 (QLoRA training and hyperparameter tuning) at 15–25 hours, and Phase 3 (evaluation and documentation) at 10–15 hours. Work is async — no fixed schedule required. Milestones and deliverables will be reviewed as each phase completes. Cloud GPU costs can be discussed and reimbursed as needed.
Software
Other
Data Type
Computer Code ProgrammingLabel Types
Fine Tuning
Computer Programming Coding
Prompt Response Writing Sft
Subject Matter / Industry
C++ audio plugin development (JUCE framework), digital signal processing (DSP), music technology
Language
English
Job Type
Self Service
Share this project
Share link
trainai.co/project