CPU High-RAM or Free T4 GPUVRAM: 8 GB
Trained Data Pipeline: Magpie Synthetic Reasoning Distillation & MinHash Corpus Scrubber
Canonical Summary (AEO): Complete training-data curation notebook: builds high-signal instruction & reasoning trace datasets (ShareGPT/ChatML), removes near-duplicates via MinHash LSH, scrubs PII/secrets, and formats DPO/GRPO preference pairs.
⚡ Reproducible Execution Steps
- Ingest raw domain documents (Markdown, PDFs, n8n workflows, or support tickets) in Cell 1.
- Run the Magpie / Teacher Synthesis pipeline in Cell 2 to generate multi-step <think> reasoning traces + verified JSON outputs.
- Run MinHash LSH (128 permutations, Jaccard threshold 0.82) in Cell 3 to eliminate near-duplicate samples.
- Run the Deterministic Validator + Rubric Scorer (1–5 scale) in Cell 4 to keep only top-quartile samples (score >= 4.2).
- Export clean ShareGPT JSONL (for SFT) and Chosen/Rejected pairs (for DPO/GRPO) to HuggingFace Hub or Google Drive.
Why Data Curation Beats Model Size (The 2026 Nemotron & DeepSeek Lesson)
Modern benchmarks ( including the NVIDIA Nemotron Reasoning Challenge and DeepSeek-R1 distillation studies) prove that 10,000 rigorously deduplicated, rubric-filtered synthetic reasoning traces outperform 500,000 noisy scraped examples when fine-tuning 7B–32B models.
4-Stage Data Curation Pipeline Included in This Notebook
- Stage 1 — Schema-Constrained Teacher Synthesis: Generates diverse user prompts and pairs them with verified
<think>...</think>reasoning chains and syntax-checked outputs (SQL, Python, n8n JSON, or clinical/legal triage). - Stage 2 — MinHash LSH Semantic Deduplication: Prevents mode collapse by stripping clusters of repetitive prompts.
- Stage 3 — Hard Secret & PII Scrubber: Uses our 8-pattern regex engine (
sk-,eyJ,Bearer,AIza, emails, phone numbers) so training corpora never memorize credentials. - Stage 4 — DPO / GRPO Preference Pair Formatter: Automatically pairs verified outputs (
chosen) against syntax-failing or low-rubric outputs (rejected) for alignment tuning.