Module FT05 — Synthetic Data Generation

Synthetic Data Generation

Where SFT and preference data actually comes from. The two dominant approaches — strong-teacher distillation and Magpie self-synthesis — and the one thesis that organizes them: the gains are in filtering, not generation.

75
minutes
8
artifacts
7
sub-sections
The dominant failure mode in synthetic-data work is not 'we generated too little' — it is 'we generated plenty and shipped the unfiltered pile.' LIMA's 1,000 curated beat Alpaca's 52,000 unfiltered. Magpie's filtering 1M to 300K improved quality more than doubling the raw. This module teaches you to generate (Magpie + teacher distillation), and flags the cleaning half (FT06) as where the gains actually live.
Key Claims
Load-Bearing Claims

The gains are in filtering, not generation. A 500-sample curated set steers better than 50,000 raw teacher outputs. The steering wheel is the curated set, not the raw pile — the data-side complement of FT00's 'steering not teaching.'

Magpie (arXiv:2406.08464, ICLR 2025) changed the defaults. Aligned instruct models generate instructions when prompted with their pre-query template alone — no seed prompts. Two passes over the same model produce a self-consistent (instruction, response) pair. Magpie-Pro-Filtered beats the official Llama-3-8B-Instruct using only self-generated data.

Strong-teacher distillation is the workhorse when you have a target prompt distribution. Real user queries → strong model → responses. Watch for style leakage (mix 2–3 teachers) and contamination (decontaminate against evals).

Evol-Instruct (WizardLM) is now a special case, not a default. Iterative LLM-driven complexification still works for hardness amplification (Pillar 4), but Magpie samples difficulty naturally and direct teacher generation is simpler when you have prompts.

distilabel (Argilla) is the standard framework for synthetic preference/SFT pipelines. Four stages: generate → judge → filter → dedup. Composes directly with FT06's cleaning machinery.

The decontamination obligation is non-negotiable. Remove eval-overlap items before training (MMLU, GSM8K, MT-Bench, your held-out eval). Plan it at generation time — don't seed with or few-shot with benchmark items. Detailed in FT06.

After This Module
01
State the thesis — the gains are in filtering, not generation — and defend it with the LIMA and Magpie ablation evidence (curation has a steeper quality curve than volume for steering).
02
Contrast strong-teacher distillation (seed prompts → strong model → responses) and Magpie self-synthesis (aligned model generates both sides from its pre-query template alone), and choose between them for a given target behavior.
03
Describe Magpie's two-pass mechanism (empty pre-query template → instruction; instruction placed properly → response) and why it works (RLHF/DPO made instruction-responding the default; the empty template surfaces the model's instruction prior).
04
Explain why Magpie outperformed Alpaca, Evol-Instruct, and UltraChat (no seed bias, self-consistency, filtering still applies), and distinguish Magpie-Air (8B) from Magpie-Pro (70B).
05
Place Evol-Instruct correctly in the history: a real advance (hardness amplification) now superseded as a default by Magpie + direct teacher generation, but still the right tool for Pillar 4 reasoning models.
06
Name distilabel (Argilla) as the standard framework and its four pipeline stages (generate → judge → filter → dedup), and explain why it composes with FT06.
07
State the decontamination obligation and why it must be planned at generation time (fix the eval set before generation; don't seed with or few-shot with benchmark items).
08
Identify the four anti-patterns — generate-without-filter, no-decontamination, single-teacher style leakage, ignored diversity — and explain how each corrupts the steering wheel.
Artifacts