Hey \ud83d\udc4b<\/p>\n
Quick project showcase. I built a dataset generation skill for Claude, Codex, and Antigravity after spending way too much on the OpenAI, Claude, and Gemini APIs.<\/p>\n
At first I was using APIs for the whole workflow. That worked, but it got expensive really fast once the work stopped being just “generate examples” and became:
generate -> inspect -> dedup -> rebalance -> verify -> audit -> re-export -> repeat<\/p>\n
So I moved the workflow into a skill and pushed as much as possible into a deterministic local pipeline.<\/p>\n
The useful part is that it is not just a synthetic dataset generator.
You can ask it to:
“generate a medical triage dataset”
“turn these URLs into a training dataset”
“use web research to build a fintech FAQ dataset”
“normalize this CSV into OpenAI JSONL”
“audit this dataset and tell me what is wrong with it”<\/p>\n
It can generate from a topic, research the topic first, collect from URLs, collect from local files\/repos, or normalize an existing dataset into one canonical pipeline.<\/p>\n
How it works:
The agent handles planning and reasoning.
The local pipeline handles normalization, verification, generation-time dedup, coverage steering, semantic review hooks, export, and auditing.<\/p>\n
What it does:
– Research-first dataset building instead of pure synthetic generation
– Canonical normalization into one internal schema
– Generation-time dedup so duplicates get rejected during the build
– Coverage checks while generating so the next batch targets missing buckets
– Semantic review via review files, not just regex-style heuristics
– Corpus audits for split leakage, context leakage, taxonomy balance, and synthetic fingerprints
– Export to OpenAI, HuggingFace, CSV, or flat JSONL
– Prompt sanitization on export so training-facing fields are safer by default while metadata stays available for analysis<\/p>\n
How it is built under the hood: <\/p>\n
SKILL.md (orchestrator)
\u251c\u2500\u2500 12 sub-skills (dataset-strategy, seed-generator, local-collector, llm-judge, dataset-auditor, …)
\u251c\u2500\u2500 8 pipeline scripts (generate.py, build_loop.py, verify.py, dedup.py, export.py, …)
\u251c\u2500\u2500 9 utility modules (canonical.py, visibility.py, coverage_plan.py, db.py, …)
\u251c\u2500\u2500 1 internal canonical schema
\u251c\u2500\u2500 3 export presets
\u2514\u2500\u2500 50 automated tests<\/p>\n
The reason I built it this way is cost.
I did not want to keep paying API prices for orchestration, cleanup, validation, and export logic that can be done locally.<\/p>\n
The second reason is control.
I wanted a workflow where I can inspect the data, keep metadata, audit the corpus, and still export a safer training artifact when needed.<\/p>\n
It started as a way to stop burning money on dataset iteration, but it ended up becoming a much cleaner dataset engineering workflow overall.<\/p>\n
If people want to try it:<\/p>\n
git clone https:\/\/github.com\/Bhanunamikaze\/AI-Dataset-Generator.git cd AI-Dataset-Generator .\/install.sh --target all --force or you can simply run curl -sSL https:\/\/raw.githubusercontent.com\/Bhanunamikaze\/ai-dataset-generator\/main\/install.sh | bash -s -- --online --target all <\/code><\/pre>\nThen restart the IDE session and ask it to build or audit a dataset.<\/p>\n
Repo:
https:\/\/github.com\/Bhanunamikaze\/AI-Dataset-Generator<\/a><\/p>\nIf anyone here is building fine-tuning or eval datasets, I would genuinely love feedback on the workflow.
\u2b50 Star it if the skill pattern feels useful
\ud83d\udc1b Open an issue if you find something broken
\ud83d\udd00 PRs are very welcome<\/p>\n<\/div>\n
submitted by \/u\/Illustrious-triffle <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
